Cognition and Amazon Web Services signed a multi-year Strategic Collaboration Agreement (SCA) on September 15, 2026, aimed squarely at one of enterprise IT's most persistent, expensive problems: legacy system modernization. The pitch is that Devin — Cognition's autonomous software engineering agent — can analyze an old system, recreate its behavior, write replacement code, and test the new system against known outputs, compressing migrations Cognition says typically take years down to months.
This is the latest in a fast-moving string of Cognition/Devin announcements this month alone — explainx.ai covered Devin's Fusion CLI cost cut on September 12 and SWE-2's frontier coding performance on September 11. The AWS deal is a different layer of the same push: distribution and enterprise infrastructure, not model capability.
TL;DR
| Question | Answer |
|---|---|
| What was announced? | A multi-year Strategic Collaboration Agreement between Cognition and AWS |
| When? | September 15, 2026 |
| What's the target use case? | Legacy system modernization — analyzing, recreating, and testing replacement code for old enterprise systems |
| Can I already buy Devin on AWS? | Yes — via AWS Marketplace, plus the Agent Toolkit for AWS usable directly inside Devin |
| Named joint customers | Mercedes-Benz, Echo Global Logistics, ActiveCampaign |
| Does Devin replace engineers? | No — Cognition frames it as engineers deciding what gets built and reviewing output, while Devin does the analysis/writing/testing |
| How does this relate to Cognition's Infosys/Cognizant deals? | Complementary — those cover consulting/integration; this covers cloud infrastructure and marketplace distribution |
What the deal actually covers
The core of the agreement is aimed at enterprise modernization projects: moving legacy workloads onto AWS faster, clearing security and engineering backlogs, and freeing internal teams to build new products rather than maintain old systems. The specific workflow Cognition describes is a three-step loop — Devin analyzes an existing system's behavior, writes replacement code, and tests the new implementation against the old system's known outputs to confirm behavioral parity before a cutover.
That's a meaningfully narrower and more concrete claim than "Devin can build anything." Legacy modernization is a well-understood, high-cost enterprise category precisely because behavioral parity is hard to verify by hand — a decades-old COBOL or mainframe system often has undocumented edge cases baked into years of patches, and manually re-implementing it risks silently dropping behavior nobody remembers is load-bearing. An agent that can systematically test new code against old outputs addresses that verification gap directly, rather than just generating code faster.
Commercially, the mechanics are already live: joint customers can purchase Devin through AWS Marketplace today, and use the Agent Toolkit for AWS directly inside Devin — meaning this SCA formalizes and deepens an integration path that already existed, rather than starting from zero. Cognition and AWS say they're "exploring deeper engineering integrations" to make Devin easier to use inside customers' existing AWS environments going forward.
Reading this against Cognition's other September moves
Cognition has been unusually active this month, and this deal fits a clear pattern when placed next to its other recent announcements:
| Announcement | Date | Layer |
|---|---|---|
| SWE-2 frontier coding model | Sept 11 | Model capability |
| Fusion CLI 39% cost cut | Sept 12 | Cost/efficiency |
| AWS Strategic Collaboration Agreement | Sept 15 | Enterprise distribution + infrastructure |
Each release addresses a different objection an enterprise buyer would raise before adopting an autonomous coding agent at scale: is it capable enough (SWE-2), is it affordable at volume (Fusion CLI), and can we actually procure and integrate it inside our existing cloud environment (AWS SCA). Layering all three in a single month suggests Cognition is deliberately building toward large enterprise modernization contracts rather than just iterating on developer-facing features.
Why legacy modernization specifically
Enterprise legacy modernization has stayed stubbornly manual for a reason: it's low-glamour, high-risk work. A mainframe or decades-old monolith often encodes business rules nobody currently at the company wrote down — quirks in tax calculations, regional compliance branches, or edge cases from a merger a decade ago. Rewriting it by hand is slow because every change needs manual verification against a system whose full behavior isn't documented anywhere except the running code itself.
That's exactly the shape of problem an agent with strong testing discipline is suited to attack: not because it's creative, but because it's mechanically exhaustive in a way humans tire of. Devin's three-step loop — analyze, recreate, test against known outputs — turns "did we preserve behavior" from a matter of manual code review into a matter of running the old and new systems side by side and diffing the results. That doesn't eliminate the need for judgment about what should change (a modernization project is also usually a chance to fix things, not just port them 1:1), but it does compress the purely mechanical verification work that has historically made these projects take years.
The bigger pattern: agents plus cloud incumbents
This deal also fits a broader 2026 trend of agentic coding companies pairing with the major cloud providers rather than staying pure independent software vendors. AWS gets a differentiated modernization offering it can sell alongside its existing migration tooling; Cognition gets distribution into AWS's enterprise sales motion and billing relationships — a channel that's genuinely hard to build from scratch as a standalone startup. Expect more of this shape of deal (agent company plus hyperscaler) across 2026 and 2027 as agentic coding tools mature from developer-facing products into enterprise procurement line items with security reviews, SLAs, and dedicated account teams attached.
What this means if you're evaluating Devin for a migration
- This is a fit for well-scoped, behavior-preserving migrations — moving a legacy system to new infrastructure while keeping its outputs identical — not open-ended greenfield development. The testing-against-known-outputs step is exactly what makes this claim credible rather than marketing-only.
- Procurement gets easier if you're already an AWS customer. Buying through AWS Marketplace means the spend can run through existing AWS commitments and billing relationships, a real practical unlock for enterprises with AWS Enterprise Discount Program agreements.
- Human review remains the stated safety net. Cognition's own framing — engineers decide what gets built and review Devin's output — matches the general pattern this blog has covered across agentic coding tools: the agent handles the mechanical, verifiable legwork; a human still owns the decision of what "correct" means for the business.
- Budget for the discovery phase separately from the migration itself. Even with an agent doing the analysis, the most time-consuming part of a legacy modernization is often figuring out what the old system is actually supposed to do — undocumented behavior, dead code paths nobody's sure are safe to remove, and tribal knowledge that left with former employees. Devin can accelerate that discovery, but it doesn't remove the need to budget time for it.
Questions worth asking before you sign up
If your organization is weighing whether to run a modernization project through this partnership, a few practical questions should come before a contract:
- What's actually being ported, and how is "parity" defined? Get specific about which behaviors count as must-preserve versus safe-to-change — a modernization project is also a natural opportunity to fix known bugs, and you don't want an agent faithfully reproducing a bug you were hoping to finally kill.
- Who reviews the generated replacement code, and against what standard? Cognition's own framing puts engineers in the review seat — confirm your organization has the in-house expertise to actually review a machine-generated reimplementation of a system nobody currently understands in full.
- What's the rollback plan if testing reveals behavioral drift late in the project? Legacy migrations fail expensively when drift is caught after cutover rather than during testing — ask specifically how Cognition's test-against-known-outputs step is structured and how much of the old system's behavior it's realistically able to capture as test cases.
FAQ
What did Cognition and AWS announce? A multi-year Strategic Collaboration Agreement, announced September 15, 2026, to deploy Devin for enterprise legacy-system modernization on AWS.
What does Devin actually do in this partnership? It analyzes an existing legacy system, recreates its behavior, writes replacement code, and tests the new system against known outputs.
Can I buy Devin through AWS today? Yes — via AWS Marketplace, with the Agent Toolkit for AWS usable directly inside Devin.
Who are the named customers? Mercedes-Benz, Echo Global Logistics, and ActiveCampaign are cited as existing joint customers.
How does this compare to Cognition's Infosys and Cognizant partnerships? Those cover consulting and system-integration services; this AWS deal covers cloud infrastructure and marketplace distribution — complementary layers of the same enterprise push.
Does this replace human software engineers for migrations? No — Cognition frames it as engineers deciding what gets built and reviewing Devin's output, with Devin handling analysis, writing, and testing underneath that review.
Related reading
- Cognition ships SWE-2: Devin's frontier coding model
- Devin's Fusion CLI cuts costs 39%
- What is an agent harness? Complete guide
- Top 10 open and closed source agent harnesses
- Marc Andreessen on Cognition, Devin, and 90% AI-written code
- TypeSafe AI's Jev: a "System One Model" for structured decisions
- Official: Cognition and AWS announcement
Details in this piece reflect Cognition and AWS's September 15, 2026 announcement. Pricing, marketplace terms, and integration depth may change as the collaboration develops.