Cognition

Software development has become a throughput bottleneck for companies that run on code. The software industry faces an escalating developer talent shortage, with demand for software engineers projected to grow by 17% from 2023 to 2033, far outpacing the 4% average across all occupations in the same period.

However, headcount growth has not translated into proportional gains in engineering output. Day-to-day engineering capacity is constrained by long-running, multi-step work: scoping tickets, navigating large codebases, resolving failures, and shepherding pull requests (PRs) through review. These coordination-heavy workflows have proven difficult to scale linearly, contributing to widespread re-insourcing, with 70% of surveyed companies bringing previously outsourced software work back in-house between 2019 and 2024.

A new wave of AI-assisted coding tools has emerged to enhance developer productivity, but most remain assistive rather than executional. While tools like GitHub Copilot have boosted task completion speeds by 55% and raised project completion rates by 8%, they primarily suggest code snippets, and human engineers still retain responsibility for planning tasks, navigating repository context, debugging failures, and coordinating reviews. As a result, productivity gains from assistive AI often shift bottlenecks downstream into testing, integration, and pull request reviews rather than removing them entirely.

These limitations have created demand for systems that can move beyond suggestion and assume responsibility for execution. With AI adoption surging, such as Copilot reaching 400K subscribers in its first month of public launch and 84% of developers using or planning to use AI in their workflows as of 2025, $8.2 billion was invested in 2024 in companies to develop autonomous end-to-end software development solutions.

Cognition is pursuing this shift with Devin, the first fully autonomous AI engineering agent designed to plan, execute, and validate multi-step software tasks. Unlike copilots that assist human programmers, Devin executes entire software projects and resolves 13.9% of real-world GitHub issues end-to-end on SWE-bench — far outperforming GPT-4’s 1.7% and Claude 2’s 4.8% as of March 2024. Cognition’s vision is to offer a scalable alternative to human engineering labor through autonomous AI coding agents that can own execution across full software workflows.

Cognition, also known as Cognition Labs, was founded in November 2023 by Scott Wu (CEO), Steven Hao (CTO), and Walden Yan (CPO). The founding team brought strong competitive programming and software engineering backgrounds.

Wu, a Harvard graduate, is a three-time gold medalist at the International Olympiad in Informatics (IOI), a 2011 Mathcounts national champion, and later the co-founder and CTO of Lunchclub from 2017 to 2022, an AI networking platform backed by Lightspeed and Coatue. He also placed third at Google Code Jam in 2021 and was named to Forbes 30 Under 30 in 2020. Hao studied computer science and math at MIT from 2014 to 2018 and became a top engineer at Scale AI after he joined the company in 2018.

Before working at Scale AI, Hao had interned at firms like D.E. Shaw and Dropbox and won a gold medal at IOI 2014, placing 6th globally. Yan, who dropped out of Harvard in 2023 to co-found Cognition, previously worked on Cursor at Anysphere from June to August 2023 and was a cofounder of DeepReason, a web3 security startup, from 2022 to 2023. He also served as managing partner at Inverted, a media consultancy that was acquired in 2021. In 2020, he earned a gold medal at IOI, placing 19th out of nearly 400 participants from over 100 countries.

The founders likely knew each other through IOI and their time at Harvard and MIT, initially exploring cryptocurrency projects before pivoting to generative AI in late 2022 as interest in models like ChatGPT accelerated across Silicon Valley. Wu said the task of building an AI to code is “deeply algorithmic” and well-suited to their background. Collectively, the team had won 10 IOI gold medals — a fact that shaped the company’s early identity and seemed to give it an edge in the field of AI programming.

The concept for Devin emerged from the team’s desire to turn the logic and process behind competitive coding into an AI system. The breakthrough came in late 2023 when the team realized that large language models like GPT-4, combined with reinforcement learning, could be trained to reason through complex, multi-step software tasks. In a pivotal moment before Christmas Day in 2023, after hours of struggling with a server issue, the team gave Devin a chance to solve it. Devin identified and deleted a faulty system file they had overlooked, successfully configuring the server. It was the first time Devin had autonomously handled a real-world task, which affirmed the company’s vision.

Cognition operated in stealth until March 2024, when it launched Devin to the public. A demo video showing Devin autonomously fixing a bug in an open-source library gained over 30 million views on X as of January 2026. The company also claimed that Devin had passed real-world engineering interviews and performed complex coding tasks without human input.

By mid-2025, Cognition had expanded from offering a single autonomous engineer to providing an enterprise-ready agentic development environment. It offers secure workspaces and execution contexts that can be deployed via SaaS or in a customer VPC, enabling teams to integrate Devin into real engineering systems. In July 2025, the company acquired Windsurf, an AI‑native IDE with built‑in agentic workflows and developer‑centric UX. The acquisition brought both product infrastructure and talent, accelerating Cognition’s ability to integrate Devin directly into a full‑featured IDE and embed its agent more deeply into day‑to‑day engineering workflows.

As of January 2026, Cognition’s core product offering centers on Devin, an autonomous AI software engineering agent, alongside an expanding suite of agent models, orchestration tools, and execution environments designed to support product-scale software development.

Devin

Devin is Cognition’s AI coding agent designed to autonomously execute engineering tasks capable of planning, coding, debugging, and integrating work into production workflows. Unlike traditional code assistants, Devin operates within an integrated development environment equipped with its own Linux shell, code editor, browser, and other standard developer tools inside a cloud-based sandbox. When given a task, whether via chat, the web platform, or a Slack command, Devin formulates a step-by-step plan, executes code, runs tests, debugs, and updates its plan as needed. Throughout the generative process, it streams updates and accepts user corrections, maintaining an interactive feedback loop.

Devin is built to handle a wide range of tasks across the software development lifecycle. It can independently build full-stack applications, find and fix bugs, learn and apply new libraries, and even train or fine-tune machine learning models. In public demos, Devin has created a dynamic web app from scratch (the Game of Life) and autonomously patched a bug in the Sympy math library. Cognition describes Devin as a “tireless, skilled teammate” that can either pair-program or complete full tasks for later review.

Devin is capable of retaining context across thousands of steps and making thousands of micro-decisions toward a goal. When asked to clean up a codebase, for example, it can devise and execute a multi-part plan involving linting, refactoring, and API updates. When faced with errors, it searches documentation or forums, and if tests fail, its built-in self-reflection loop allows it to adjust strategy and retry. Other features include voice command input and enhanced repository context for working with large codebases. Devin is available as a web app, Slack bot, or VS Code extension and integrates with tools like GitHub to act as a junior developer on a team.

Source: Cognition

In April 2025, Cognition released Devin 2.0, which introduced several enterprise-grade enhancements aimed at improving transparency and collaboration in complex workflows. Devin 2.0 can perform proactive codebase exploration and generate a detailed execution plan that users can edit or approve before task execution begins, a feature which Cognition calls interactive planning. The update also introduced Devin Search, which allows users to query their codebase directly and receive responses backed by cited code, improving visibility and codebase comprehension. Following the release of Devin 2.0, Cognition improved the platform to Devin 2.1 in May 2025, focused on providing greater clarity with confidence ratings for its Linear and Jira integrations, and better success in large codebases with better context. Users are able to ask any questions about the code, and Devin will provide a response informed by code snippets. Additionally, Devin Wiki automatically indexes repositories every few hours, producing browsable documentation and architecture diagrams that link directly to relevant parts of the code.

In January 2026, Cognition also released Devin Review, a code review tool that works on any public or private GitHub PR. Devin Review is designed to automate large portions of the code review process by analyzing diffs, reasoning over repository context, and generating structured review comments that flag bugs. Devin Review reasons across files and code history, and also offers an inline interactive chat enabling users to ask about changes.

Source: Cognition

At the enterprise level, Cognition offers advanced capabilities through MultiDevin, a version of Devin designed for large-scale, parallelized task execution. MultiDevin consists of one “manager” Devin that coordinates up to ten “worker” Devins, each assigned a small, isolated subtask. Worker Devins run in parallel, and their successful outputs are automatically merged into a single codebase. This setup is particularly effective for workloads that involve a high volume of repetitive subtasks, subtasks of junior engineer-level difficulty, and those that are isolated, incremental, and objectively verifiable. This may include tasks such as linting, refactoring, or updating API calls across a large monolithic codebase. Devin is best suited for projects with minimal interdependencies, where each component of the work can be completed and validated independently.

To support secure deployment at scale, Cognition offers a Virtual Private Cloud (VPC) deployment model. The architecture of the VPC system separates the Devin Backend, composed of Devin’s Brain, event metadata, and backend servers, from the Customer VPC, which hosts Devin’s Dev Box. The Dev Box contains a Linux shell, code editor, browser, and the Devin Agent itself, all operating inside a customizable environment. Customer data resides fully within their VPC, and communication between systems is encrypted in transit. Devin integrates with developer ecosystems, including GitHub, GitLab, and Snowflake, as well as customer-provided SaaS tools like Slack, Stripe, Teams, and Zapier, through secure API tokens. This infrastructure enables Devin to act as a safe, embedded AI teammate, capable of scaling engineering workflows while maintaining full enterprise compliance and control.

Source: Cognition

In Cognition’s November 2025 performance review, Cognition reported that Devin has become significantly more efficient and capable in real engineering workflows: up to 4× faster problem solving, 2× more resource efficient, and with ~67% of its PRs merged versus ~34% previously.

Source: Cognition

Additionally, Devin is designed to resolve vulnerabilities flagged by static analysis tools (e.g., SonarQube, Veracode), saving organizations 5-10% of total developer time, and also increasing test coverage from 50-60% to 80-90%. In production, companies like Nubank have used Devin to automate large-scale refactoring of monolithic codebases, resulting in an 8× improvement in engineering efficiency and 20× cost savings as of January 2026.

Source: Cognition

Model Layer

While early versions of Devin relied on frontier foundation models like GPT-4 Turbo, Cognition has introduced proprietary agent models optimized specifically for software engineering tasks. This includes the SWE-1.5 model families, which are designed to balance reasoning, execution speed, and cost efficiency for long-horizon coding tasks. Popular use cases of SWE-1.5 include deeply understanding large codebases, building end-to-end full-stack apps, and easily editing configurations without needing to memorize field names. Cognition’s goal for SWE-1.5 was to create the fastest coding agent experience available, allowing for frontier-level performance at 13x the speed of Claude’s Sonnet 4.5.

Source: Cognition

Cognition has also released specialized models like SWE-grep and SWE-grep-mini, which focus on fast, parallel repository search and context retrieval. SWE-grep and SWE-grep-mini match the retrieval capabilities of frontier coding models by using parallel tool calls, limited serial turns, and an inference that is 20x faster and 4.5x faster than Claude’s Haiku 4.5 at 140 tokens/second. The company worked with Cerebras, the fastest inference provider, to deploy and optimize the SWE-1.5 and SWE-grep family.

In July 2025, Cognition signed a definitive agreement to acquire Windsurf, an “agentic IDE” designed to support long-running AI coding agents inside a native development environment. Following the acquisition, Cognition’s proprietary agent models, including SWE-1.5 and SWE-grep, became available directly inside Windsurf, allowing developers to interact with these models through a native IDE. In this setup, Windsurf serves as a tightly integrated execution and control surface where agent models can reason over large codebases, run tools, and apply changes with lower latency and faster iteration loops.

Customer

As of January 2026, Cognition’s customers included OpenSea, Ramp*, Nubank, Lumos, Microsoft, and Curai Health. The company has further highlighted Goldman Sachs, Citi, Dell, Cisco, Ramp, Palantir, Nubank, and Mercado Libre as users of Devin and associated products following the Windsurf acquisition and expanded go-to-market efforts. These early customers have primarily been engineering-centric, growth-focused, mid to large tech companies, many of which maintain substantial codebases and developer teams. They demonstrate a clear emphasis on developer productivity and willingness to experiment with emerging technologies like AI coding agents. For example, Microsoft highlights Devin in its Azure customer stories as a tool being used by engineers to scale productivity and accelerate modernization efforts.

Several industry verticals stand out for Cognition’s ideal customer. In fintech, companies like Ramp and Nubank are adopting Devin to modernize legacy systems, where technical debt often grows faster than it can be resolved. In these environments, Devin’s ability to automate refactors, migrate APIs, and run verification loops provides meaningful leverage. Technology startups such as Lumos, which tend to operate with lean engineering teams, are another natural fit, as they leverage AI tools like Devin to accelerate development cycles and focus human engineers on higher-order logic. Cognition’s other customers reflect a cross-industry interest in Devin’s capabilities, such as OpenSea in Web3 and Curai Health in healthcare — sectors where software quality and iteration speed can be key differentiators.

Cognition’s 2025 reporting suggests that Devin had become operational in engineering teams at thousands of companies, with the agent having merged hundreds of thousands of pull requests across these deployments. Cognition’s enterprise focus, including partnerships with Infosys and broader go-to-market alignment post-Windsurf, suggests the company is targeting larger enterprises, like Fortune 500 customers, that have both the scale and complexity of engineering work to benefit most from agent automation.

Market Size

Cognition operates within the emerging generative AI coding assistants market, which was valued at $25.9 million in 2024 and is expected to grow to $97.9 million by 2030, reflecting a CAGR of 25.5% from 2024 to 2030. While current market size figures are relatively small, the gen AI coding assistants market is only beginning to monetize and has the potential to grow significantly as large enterprises increasingly adopt agents to automate repetitive, tedious software engineering tasks.

However, broader research into AI code tools and assistants has led to estimated market growth from $4.9 billion in 2023 to $26 billion by 2030 at 27.1% CAGR. This market can be segmented into three layers: code autocompletion tools (e.g., Copilot), AI code review and optimization tools, and a new third category — fully autonomous coding agents. Cognition targets this last segment, which is both the most nascent and potentially the most disruptive.

Beyond direct software tooling revenue, Cognition’s long-term opportunity is often framed against the broader economic value of software development itself. As of 2025, there were approximately 30 million software developers worldwide, a number expected to reach 45 million by 2030. Assuming each developer generates $100K per year in economic value, Andreessen Horowitz estimates that deployment of AI can at least double developer productivity, resulting in a projected $3 trillion annual impact from AI-augmented software development. As more of software engineers’ workflows become automatable, Cognition’s positioning within this market allows it to capture value well beyond the current coding assistant segment.

The competitive landscape for AI coding assistants can be divided between established incumbents and a wave of well-funded startups in categories like AI code assistants, agentic developer tools, AI-native code editors, and foundation models for software engineering. Cognition differentiates itself by targeting fully autonomous software engineering agents that can plan, execute, and validate multi-step tasks, rather than merely assisting developers within an existing workflow.

Incumbents

GitHub Copilot: GitHub Copilot, launched in June 2021, offers AI-powered code suggestions and in-IDE autocompletions through its integration with editors like VS Code and JetBrains. Now part of the broader Copilot X platform, Copilot supports chat-based code explanation and has over 20 million all-time users, 8 million paid subscribers, and 50K business team subscribers. Backed by Microsoft and OpenAI, who together have raised nearly $18 billion, Copilot remains widely adopted. Despite its scale, Copilot remains fundamentally assistive. It does not autonomously plan or execute full engineering tasks, manage long-running workflows, or recover from failure without human intervention.

Amazon Q: Amazon’s Q Developer, launched in November 2023, offers a more comprehensive feature set, including code completions, chat-based assistance, security scans, and refactoring. Integrated directly into the AWS ecosystem, it claims to improve coding speeds by 80% and boost productivity by 40%, with an acceptance rate of 37%. While Q Developer shares more workflow automation overlap with Devin than Copilot, Devin still surpasses it in long-term planning, recovery from failure, and execution of complex, multi-step engineering goals.

Claude Code: Claude Code, launched in February 2025 by Anthropic, is an agentic coding tool embedded directly in the terminal. It enables developers to edit files, run and fix tests, resolve merge conflicts, and query architectural logic via natural language. In benchmark testing, Claude Code achieved a 70.3% score on SWE-bench Verified and has been shown to save over 45 minutes per task in enterprise workflows, reducing test run time by 95%. As of January 2026, it was used by teams at top technology firms like Cursor, Replit, and GitHub. However, Claude Code operates without a dedicated, persistent development environment and remains human-steered rather than autonomous.

GPT-4: Finally, GPT-4, accessed through ChatGPT or API integrations, is often used in coding contexts via custom tools like Auto-GPT or SWE-agent. While GPT-4 ranks high on coding benchmarks like HumanEval, it lacks structured memory and planning and typically requires significant orchestration to complete multi-step development tasks. In one study, GPT-4 scored only ~1% on SWE-bench when deployed without fine-tuning. Devin, by contrast, is purpose-built with long-term memory, a structured planning loop, and a sandboxed development environment, enabling it to reliably complete tasks that general-purpose models like GPT-4 cannot handle alone.

AI Code Editors

Cursor (Anysphere): Among AI-native code editors, Cursor (owned by Anysphere) has emerged as a leading competitor to Devin. Released in May 2023 by Anysphere, Cursor is built from the ground up as an AI-first IDE. It supports natural language commands, smart autocompletion, chat-based assistance, and multi-file edits powered by a custom-trained model. Anysphere, founded by MIT alumni in December 2022, has attracted massive investor interest, raising a $2.3 billion Series D funding round at a $29.3 billion valuation in November 2025. As of January 2026, Cursor had over 1 million users, with 360K paying customers, and has surpassed $500 million in ARR. While Cursor excels at improving developer efficiency inside the IDE, Devin differentiates itself by being an autonomous agent, able to plan and execute full engineering workflows across repositories without constant human input.

Augment Code: A newer entrant is Augment Code, which launched from stealth in April 2024 with $252 million in Series B funding at a $977 million valuation. Founded by ex-Google and Microsoft engineers Igor Ostrovsky and Guy Gur-Ari in 2022, Augment is focused on helping engineering teams automate backlog reduction and improve software quality at scale. Its product is designed to turn structured inputs like Jira tickets into complete merge requests, with initial usage across “hundreds” of developers in early access, including payment startup Keeta, according to Ostrovsky in April 2024. While Devin has focused on arbitrary user-prompted tasks and full-cycle development, Augment leans into workflow automation and quality assurance across enterprise codebases — suggesting a complementary, though distinct, strategic focus.

Foundation Models for SWE

Poolside AI: Poolside AI, founded in 2023 by Jason Warner (former CTO of GitHub) and Eiso Kant, is developing foundation models from the ground up for code-related applications such as autocomplete, intelligent code search, and codebase navigation. In October 2024, the company raised a $500 million Series B round from investors, including eBay and NVIDIA, bringing its total funding to $626 million and valuing the company at $3 billion. The funding enabled the company to bring 10K NVIDIA GPUs online for training in October 2024. In October 2025, Nvidia was reportedly investing up to $1 billion in the company as a part of a $2 billion funding round that Poolside is raising at a $12 billion valuation.

Poolside offers an enterprise coding assistant powered by its proprietary models, Malibu and Point, for code generation, testing, documentation, and real-time code completion. Poolside’s differentiation focuses on being model-first, training proprietary foundation models for software and deploying them inside enterprise boundaries, whereas Cognition focuses on an agent-first systems approach oriented around autonomous execution, planning, and tool use in Devin.

Magic AI: Magic AI is also building proprietary foundation models for software development. Founded in March 2022, the company secured a $320 million investment in August 2024 from investors like Eric Schmidt and Atlassian at a valuation of approximately $1.5 billion, bringing its total funding to $626 million as of February 2026. Magic partnered with Google Cloud to build dedicated AI infrastructure, including systems equipped with 8K NVIDIA H100 GPUs and upcoming Blackwell chips. The company has stated that its long-term objective is to scale compute to 160 exaflops for training.

Technically, Magic has introduced a long-context architecture called Long-Term Memory (LTM), which reportedly supports 100 million token windows — significantly larger than the 2 million supported by other models such as Gemini as of May 2024. This capability could allow Magic’s models to process and reason over entire enterprise-scale codebases without needing to break problems into smaller parts. While Devin focuses on orchestration, planning, and execution using existing model backends, Magic is pursuing scale and context size as differentiators in its model architecture. Both approaches address different components of the software development stack and may converge over time.

AI Code Agents

Factory: Founded in 2023, Factory offers an agentic development platform built around autonomous “AI droids” that assist throughout the software engineering lifecycle. These agents integrate with version control systems and issue trackers to pick up tasks, implement code changes, open pull requests, and handle feedback during code review. The company raised a $50 million Series B in September 2025, led by NEA, Sequoia Capital, J.P. Morgan, and Nvidia, which valued the startup at $300 million. The company has raised a total of $70 million as of February 2026. Factory emphasizes fast iteration, working closely with the LangChain ecosystem to refine prompting strategies and response quality.

Factory’s platform for Agent-Native Development was globally rolled out at enterprise engineering organizations, including MongoDB, EY, Bayer, Zapier, and Clari. These organizations are seeing 31x faster feature delivery, 96.1% shorter migration times, 95.8% reduction in on-call resolution times, allowing developers to focus on design and architecture. On the SWE-Bench benchmark, in June 2024, Factory’s droid solved approximately 19% of tasks — higher than many baseline models but lower than Devin’s reported scores. Compared to Devin, Factory appears more focused on team augmentation rather than full autonomy, which may appeal to organizations that prefer keeping engineers in the loop.

Genie (Cosine): Genie, developed by Cosine (YC W23) in 2023, is another autonomous coding agent that has gained attention primarily through performance benchmarks. Despite raising only a $2.5 million seed round in August 2024, led by Soma and Uphonest Capital, Genie achieved 30% task completion on SWE-Bench — the highest reported by any agent, as of August 2024. This marks a 57% relative improvement over Factory’s 19% and significantly exceeds other published results. Cosine attributes this to its approach of encoding human-like reasoning patterns into the model, aiming to emulate how experienced engineers structure and solve problems. Genie is built on top of OpenAI’s long-context model, with additional fine-tuning and behavior-layer techniques developed in-house. The company has raised a total of $3.5 million in funding as of February 2026.

As of February 2026, Cognition offers Devin to customers with a SaaS business model, offering subscription-based access to its AI coding agent. After an invite-only beta period, the company launched its official pricing in December 2024 when Devin became generally available. The pricing is split into two main tiers: a Team Plan at $500 per month and a customizable Enterprise Plan.

The $500 per month Team Plan includes full access for an entire engineering team, with no per-seat limit. This plan provides the use of the standard version of Devin, allowing for autonomous task execution, collaboration via natural language, and learning from historical interaction patterns. The pricing is based on compute usage, or “ACUs,” rather than the number of users. If usage exceeds the included compute quota, teams can opt into usage-based billing on top of the base subscription. Team users also gain access to early feature releases and research previews.

The Enterprise Plan offers access to Devin Enterprise, which includes more advanced capabilities such as Custom Devins (fine-tuned versions for specific workflows) and advanced mode, where Devin can analyze sessions, start batch sessions, and manage organizations’ knowledge base. Additional enterprise features include virtual private cloud (VPC) deployment, a dedicated account team, and custom legal and security terms.

By offering a flat-rate base price for teams and usage-based scalability, Cognition lowers the barrier to initial adoption while creating a path to monetize increased usage. Over time, this structure allows the company to expand revenue through higher-volume usage, enterprise customization, and secure deployments, particularly for organizations with complex codebases.

Devin has generated interest within the developer and broader tech community since its public debut. The initial demo, published by Cognition in March 2024, received over 30 million views shortly after its release. While Cognition has not released official user numbers, the product has likely reached hundreds or thousands of developers across its customer base within the first few months of launch, given the size of its early enterprise customers. Independent developer forums discussing Devin’s technical capabilities have emerged, and open-source projects such as OpenHands have begun replicating its architecture, indicating early influence on the engineering community.

Devin has already contributed measurable engineering output to customer deployments. As of February 2025, at Linktree, Devin authored approximately 300 pull requests over one month, of which about 100 were successfully merged. These contributions included bug fixes, small features, and initial implementations of more complex functionality. One example was the integration of newer social platforms, such as RedNote and Lemon8, which Devin handled end-to-end, modifying backend systems, updating URL handling, and adjusting UI components. This helped Linktree streamline internal processes that traditionally required extensive planning and developer coordination.

Cognition has disclosed rapid revenue growth following Devin’s commercial launch, reporting “continuous revenue increases.” Before acquiring Windsurf, Cognition's Devin ARR grew from $1 million ARR in September 2024 to $73 million ARR in June 2025, as usage increased substantially.

The company has not disclosed total user counts, but Devin is available to all engineering teams as of February 2026. Given the team-based licensing model (no per-seat limits) and several customers with large engineering teams, it is likely that thousands of developers have used Devin directly as of February 2026. With ARR reportedly exceeding $73 million post-Windsurf acquisition, Cognition appears to be balancing aggressive revenue growth with continued product iteration.

Following Cognition’s acquisition of Windsurf in July 2025, Devin and Cognition’s proprietary agent models also gained a broader distribution surface through a native IDE environment, extending access to developer teams already using AI-assisted coding tools. Cognition has also signaled international expansion, partnering with Infosys to deploy Devin across Infosys’ internal engineering operations and global client engagements.

Devin has additionally attracted validation from prominent figures in the tech ecosystem. Eric Glyman, co-founder and CEO of Ramp, described Devin’s demonstration as the “single most impressive demo” he had seen in the past decade. Aravind Srinivas, CEO of Perplexity, commented that Devin was the first AI agent he had seen “that seems to cross the threshold of what is human level and works reliably.”

Cognition has raised substantial venture funding since launching Devin. In September 2025, Cognition raised a $400 million at a $10.2 billion valuation led by Founders Fund. Other investors include 8VC, Neo, Elad Gil, Definition Capital, and Swish VC. The round brought the company’s total funding to $896 million.

The company has seen rapid growth in its funding and valuation. In March 2024, the company secured a $21 million Series A round led by Founders Fund at a valuation of $350 million. The following month, in April 2024, Founders Fund led a significantly larger round of $175 million, increasing Cognition’s valuation to approximately $2 billion. Other participating investors include 8VC, Elad Gil, Conviction Partners, and Khosla Ventures.

Expand Key Partnerships

Cognition has the opportunity to expand through a combination of strategic partnerships, investor tailwinds, and early-mover advantage in the autonomous AI agent space. A key early milestone was its partnership with Microsoft, announced in May 2024. The collaboration integrated Devin into Microsoft’s developer ecosystem, with a focus on increasing productivity in areas such as code migration and modernization. This relationship increased Cognition’s exposure among enterprise software teams and developers using tools like Visual Studio Code and GitHub while also enhancing credibility with potential customers and investors.

Since 2025, Cognition has also broadened its focus to distribution-level partnerships, collaborating with Infosys, a global digital services and consulting company, to deploy Devin across the company’s internal engineering operations and global enterprise clients. As the first “large digital services and consulting firm to deploy agentic tools at this scale,” Cognition gained a potentially powerful channel for large-scale validation.

Industry and Investor Tailwinds

The broader market environment has also been favorable. In 2024, venture capital interest in AI developer tools accelerated. Large funding rounds included $10 billion raised by Databricks and $260 million raised by Glean Technologies, reflecting sustained demand for enterprise AI solutions. Overall, AI represented more than 60% of all venture funding in Q4 2024. This momentum has continued in 2025, with investors increasingly favoring agentic systems over single-function copilots. As enterprises move from experimentation to production deployment, demand is shifting toward AI systems that can execute multi-step tasks autonomously, integrate with internal tooling, and operate within security and compliance constraints. This trend aligns closely with Cognition’s product direction, strengthening its ability to secure further capital, form strategic partnerships, or pursue acquisitions to enhance Devin’s capabilities.

Early-Mover Advantage

Cognition benefits from being among the first to market with an autonomous software engineering agent. Investment in AI companies rose by over 80% year-over-year, from $55.6 billion in 2023 to over $100 billion in 2024. TechCrunch further reported that total AI startup funding reached $110 billion in 2024 — a 62% increase despite broader declines in venture funding. As of January 2026, in the segment Devin operates in, the global market for generative AI coding assistants is projected to grow from $25.9 million in 2024 to $97.9 million by 2030, representing a 24.8% CAGR from 2024 to 2030. Investor interest in “fully autonomous agents” is increasing, which aligns with Cognition’s positioning.

Beyond timing, Cognition’s early-mover advantage lies in its agentic depth, and newer features like Devin Review extend this capability across the software development lifecycle. Furthermore, Devin has the potential to expand beyond “junior-level work.” As Devin accumulates repository-level context and historical execution data, Cognition may be able to expand from task execution into higher-order engineering functions such as system refactoring, migration planning, and architectural review, and become one of the first agents to conduct this type of work.

Exaggerated Capabilities

Cognition has marketed Devin as a breakthrough in autonomous software engineering, but the company has faced scrutiny regarding the accuracy of its product demonstrations. One example involved a detailed review in April 2024 by the YouTube channel “Internet of Bugs,” which analyzed an Upwork task featured in a demo. The video alleges that Devin may have operated on pre-configured files with known bugs, calling into question the authenticity of its problem-solving process. These concerns have gained traction in online communities such as Reddit, where users have expressed skepticism about the agent’s capabilities and the framing of its performance.

Additional independent testing has further highlighted gaps between demonstration and real-world functionality. An evaluation by Answer.AI found that out of 20 tasks assigned to Devin, only three were successful, while 14 failed and three were inconclusive, suggesting a 15% success rate. The report also noted that Devin sometimes took days to complete tasks that typically take human engineers a few hours. Although Cognition has responded to some of these criticisms, maintaining transparency about limitations and performance will be important for preserving trust within the developer community.

Intense Competition

The autonomous coding space is attracting attention from well-capitalized incumbents and startups alike. Large technology companies such as Microsoft (via GitHub Copilot), Amazon (AWS Q), and Anthropic (Claude Code) are building or expanding competing solutions. Magic AI and Poolside AI are developing software-specific foundation models with significant funding and infrastructure. Meanwhile, AI-native IDEs like Anysphere (Cursor) have demonstrated rapid adoption with 1 million users (360K paying) within 16 months of launch. If a competitor integrates a more advanced model with strong planning and memory into a full-stack coding agent, it could reduce Cognition’s differentiation and capture market share.

Given Cognition’s relatively small team and limited operational history, it may face challenges defending its position if larger players offer comparable autonomous capabilities as part of broader developer ecosystems.

Scaling Challenges

Devin’s architecture involves long-context inference, multi-step planning, and sandboxed development environments — all of which can be compute-intensive. While the $500 per month team plan makes Devin accessible to smaller organizations, it may not fully cover usage costs for teams running high-volume or enterprise-scale workloads. Without careful cost control, heavy usage could compress margins or lead to higher-than-expected operational expenses.

If revenue growth lags behind infrastructure costs, Cognition may be forced to raise additional capital under less favorable terms or limit access to high-compute functionality — both of which could slow product adoption. Scalability and pricing optimization will be key as the company transitions from early-stage adoption to broader enterprise deployment.

Organizational Friction

Even when AI tools deliver individual productivity gains, these benefits can be neutralized by broader organizational friction. While 68% of developers reported saving over 10 hours with AI tools according to one 2025 survey, organizational inefficiencies still cause developers to lose significant time due to poor communication, unclear project direction, and difficulty retrieving information.

Developers reported losing at least 6 hours weekly to inefficiencies according to the same survey, and 50% lost over 10 hours due to organizational friction. Empirical studies further indicate that adoption barriers, including knowledge gaps, governance misalignment, and workflow integration challenges, limit the extent to which organizations can realize AI’s potential. As a result, Cognition’s success depends not only on agent performance, but on enterprises’ willingness and ability to redesign engineering workflows around machine ownership of execution. This may require Cognition to invest in change-management tooling, including clearer task specification, auditability, and finding an appropriate balance between AI and human tooling.

Cognition is the maker of Devin, an AI coding agent capable of planning and executing complex engineering tasks end-to-end in a sandboxed environment. Devin can write code, debug, deploy applications, and even learn new tools independently, functioning as a versatile teammate across the software lifecycle. Despite intense competition from both incumbent players and specialized startups, Cognition’s early focus on agentic depth, expanding enterprise deployments, and strategic partnerships allows the company to maintain an early lead in the rapidly expanding AI coding market.

*Contrary is an investor in Ramp through one or more affiliates.