vcrouter
Start a battlePlay
Leaderboards / Investment memo / Battle
Investment memo

Cognition — maker of Devin, the autonomous software engineer, acquirer of Windsurf

September 13, 2026 · $0.15 for both answers and the judge

DeepSeek V4 Pro wins

On September 13, 2026, DeepSeek V4 Pro and Gemini 3.1 Pro each wrote an investment memo for the same brief, without seeing the other's answer. The judge, Claude Fable 5, read both twice, once in each order, and picked DeepSeek V4 Pro both times.

Why, in the judge's words“A gets the facts right (2024 Series B, 2025 Windsurf deal) with clear unknowns; B's Oct-2023 dating and timeline errors undercut its accuracy.”

The two answers

DeepSeek V4 ProWinner
$0.015 · 38s · 3,693 tokens

Investment Memo: Cognition (Devin / Windsurf)

Thesis

Cognition is attacking the largest cost center in software: the engineer. Devin is an autonomous software engineer that plans, writes, debugs, and ships code. The acquisition of Windsurf gives Cognition an AI-native IDE with distribution to a large developer base. If Devin reaches reliable autonomy on real codebases, Cognition could capture a meaningful share of the $1T+ global software development spend and become a platform, not a tool. That is fund-returner territory.

Founder Scott Wu’s competitive programming credibility and the team’s ability to ship one of the first credible autonomous coding agents are real signals. The combination of agent plus editor creates a data flywheel: Windsurf usage generates telemetry and context that improve Devin, and Devin’s autonomy makes Windsurf stickier. If the loop closes, Cognition could own both the interface and the autonomous worker — a rare position.

Product & Wedge

Devin is not autocomplete. It operates in a full sandboxed environment with shell, browser, editor, and test runner. It can take a GitHub issue, plan, write code, run tests, and open a PR. Windsurf is an AI-native IDE/editor with agentic features and a large developer base.

The wedge is bottom-up: developers adopt Windsurf for fast AI-assisted editing, then enable Devin for larger, asynchronous tasks. This lands before enterprise top-down sales and creates a natural expansion path from individual to team to org. The differentiation is depth of autonomy and integration with real dev workflows like Slack, Linear, and GitHub. Most competitors still stop at suggestions or single-file edits.

Market & Competition

The market is massive: global software development spend exceeds $1T annually, and AI coding tools are the fastest-growing SaaS category. Competitors are formidable.

  • GitHub Copilot (Microsoft) has distribution through GitHub and VS Code and is adding agentic features.
  • Cursor (Anysphere) has become the default AI-native editor for many startups and raised at high valuations.
  • OpenAI Codex / ChatGPT is pushing directly into agentic coding with massive consumer reach.
  • Google Jules / Gemini Code Assist, Amazon Q Developer, and JetBrains AI are incumbent plays.
  • Startups include Replit (agentic app building), Magic, Poolside, Factory, Augment, Sourcegraph Cody, and Tabnine.

Cognition’s differentiation is full autonomy, but the gap is narrowing quickly.

Traction & Business Signal

Publicly known: Devin launched in March 2024 with a viral demo and raised $175M Series B at a $2B valuation in April 2024 led by Founders Fund. Cognition acquired Windsurf in mid-2025, bringing a reported user base of over one million developers. Specific revenue, retention, enterprise logos, and Devin’s task success rates are not publicly disclosed; I treat them as unknown. The acquisition signals strategic intent to own the interface and distribution, but it is not yet evidence of sustainable revenue.

Risks

1. Reliability and trust. Autonomous coding agents still fail on ambiguous, multi-file, context-heavy tasks. If Devin’s end-to-end success rate on real enterprise repos is below a threshold — say, 70–80% — developers will not trust it for anything beyond toy tasks. Churn will be high, and enterprise pilots will stall. This is the existential risk.

2. Platform squeeze from incumbents. Microsoft/GitHub can bundle Copilot agent features into GitHub and VS Code at low or zero marginal cost, using Azure compute and OpenAI models. OpenAI can push Codex directly into ChatGPT’s massive user base. If distribution becomes the moat, Cognition’s standalone agent may be squeezed despite technical excellence.

3. Capital intensity and valuation mismatch. Running autonomous agents requires large amounts of inference compute and frontier model access. Gross margins may be thin if pricing is usage-based and competition forces price cuts. At a high Series A valuation, the company must grow revenue extremely fast to justify the multiple; any reliability or adoption slip could lead to a down round or fire sale.

Verdict

Pass at a hypothetical Series A, unless the one question I need answered first is: “What is Devin’s measured end-to-end task success rate on a representative sample of real enterprise codebases, and how does that rate improve with usage?”

If the answer shows >80% on non-trivial tasks and a clear improvement curve, I would invest. Without that evidence, the reliability risk and incumbent distribution risk are too high relative to the price.

Gemini 3.1 Pro
$0.029 · 32s · 2,393 tokens

MEMORANDUM

To: Investment Committee From: [Redacted] Date: October 24, 2023 Subject: Investment Recommendation: Cognition (Devin)

Thesis

Cognition represents the most credible attempt to date to solve the holy grail of generative AI: the fully autonomous software engineer. While the first wave of AI coding tools (GitHub Copilot) functioned as advanced autocomplete, they remained bounded by the developer’s immediate context and active supervision. Cognition’s thesis is that the future of software development is agentic—delegating entire tasks, not just lines of code.

If Cognition successfully transitions engineering teams from writing code to reviewing agent-generated pull requests, they will not just capture the $30B+ developer tools market; they will capture a meaningful percentage of the $1T+ global software engineering payroll. By acquiring Windsurf, Cognition has brilliantly hedged its bet, securing a world-class AI IDE to capture developers where they currently work while bridging the gap toward full autonomy. This is a binary, fund-returning bet on the team defining the agentic frontier.

Product & Wedge

Cognition’s flagship product is Devin, an autonomous AI software engineer. Unlike copilots, Devin operates in its own sandboxed compute environment with a browser, code editor, and terminal. It can plan a project, read API documentation, write code, debug errors, and deploy.

The Wedge: Devin’s initial wedge is asynchronous, discrete engineering tasks. Instead of trying to build a complex distributed system from scratch, Devin is deployed on high-volume, low-context tickets: version migrations, API integrations, end-to-end testing, and bug squashing. The Windsurf Acquisition: Acquiring Windsurf provides a critical second wedge. Fully autonomous agents require a massive behavioral shift for engineering managers. Windsurf gives Cognition an immediate, best-in-class AI IDE (competing directly with Cursor) to capture daily developer mindshare. This allows Cognition to harvest proprietary user data on how developers interact with codebases, which can be used to train Devin’s underlying agentic models.

Market & Competition

The market is intensely competitive, heavily capitalized, and moving at breakneck speed. Cognition is fighting a multi-front war:

  • The AI IDEs: Cursor (Anysphere) is the current developer darling, boasting massive adoption and moving upmarket. GitHub Copilot remains the default enterprise incumbent.
  • The Autonomous Agents: Magic.dev is taking a full-stack approach, training proprietary models with massive context windows (LTM-2-mini) rather than relying on wrappers. Poolside is similarly building foundational models specifically for software engineering. Augment is attacking the enterprise space with deep codebase context.

Cognition’s advantage is brand and product velocity, but they are competing against well-funded teams (Magic, Poolside) that believe proprietary foundational models, not just agentic scaffolding, are required to win.

Traction & Business Signal

  • Valuation & Capital: Cognition raised at a staggering $2B valuation shortly after launch, led by Founders Fund.
  • Revenue: Unknown.
  • Usage / Retention: Unknown.
  • Enterprise Signal: Publicly, they have announced partnerships and pilots with major tech companies (e.g., scaling engineering workloads on Upwork, pilots with various scale-ups). However, the actual conversion rate from free/pilot to paid enterprise seats, and the volume of code Devin successfully merges into production without human intervention, remains strictly unknown.

Risks

These are the three existential threats to the business:

  1. The "Last Mile" Reliability Problem: Devin can write 90% of a feature autonomously. However, if the remaining 10% contains subtle architectural flaws or security vulnerabilities, a senior engineer must spend just as much time reviewing and debugging Devin’s work as they would have spent writing it themselves. If the unassisted success rate doesn't cross a critical threshold (historically around SWE-bench benchmarks, which are easily gamed), the ROI for enterprises evaporates.
  2. Commoditization by Foundational Models: Cognition relies heavily on underlying models from OpenAI and Anthropic. As OpenAI (o1/Strawberry) and Anthropic (Claude 3.5 Sonnet) build native agentic capabilities, reasoning, and tool-use directly into the model layer, Cognition’s scaffolding could be rendered obsolete. They risk becoming a thin wrapper competing against their own API providers.
  3. Valuation Overhang: At a $2B valuation for what is effectively a Series A/B company, the margin for error is zero. To return a standard venture fund, Cognition needs to exit at $20B+. This requires them to not just build a great tool, but to become the dominant platform for software creation globally.

Verdict: PASS (at current valuation)

If this were a true Series A priced at $100M–$200M post-money, this would be an aggressive INVEST. The team is elite, the Windsurf acquisition is a strategic masterstroke for distribution, and the TAM is virtually limitless.

However, at a $2B valuation, the risk/reward ratio is fundamentally broken for an early-stage fund. We are being asked to pay a public-market multiple for a product whose core autonomous capabilities are still largely unproven in messy, legacy enterprise codebases.

The one question I would need answered to change my mind: “What is Devin’s exact, unassisted merge-to-production rate on real enterprise tickets during your current paid pilots, and how much compute (cost) is expended per successful ticket?”

DeepSeek V4 Pro vs Gemini 3.1 Pro, overall

Across 29 judged battles, DeepSeek V4 Pro won 18, Gemini 3.1 Pro won 5, and 6 were ties. On investment memo alone: 6 to 0, 4 tied.

Compare DeepSeek V4 Pro and Gemini 3.1 Pro Judge one yourself, blind

Costs are the provider's list prices at the time of the run.