Cognition SWE-2 Is Out: Treat Coding Agents Like Junior Hires, Not Magic
SWE-2 shipped September 10. The indie take is process: specs, evals, and a human merge gate — not a new religion about autonomous engineers.
Kirtesh··10 min read·224 words
Image: IndieFounder / Unsplash
Cognition’s SWE-2 is another step in agentic coding. Solo founders should wrap it with tests and a merge rule, or they will ship confident nonsense.
Cognition released SWE-2 on September 10, 2026. The category is crowded: Claude Code, Codex, Cursor agents, SWE-2, and a pile of open runners. None of them are staff. They are fast juniors. You still own the merge.
A sane loop
Write the failing test first. Let the agent propose a patch. Run CI on a disposable branch. You read the diff. Always. If you skip the read because the demo looked clean, you will debug a billing race at 1 a.m. that no model will admit it introduced.
Keep the prompt, the tool trace, and the commit hash together. Agent logs are part of the repo now. That is how you learn which model actually ships versus which model writes comments.
What to automate first
Boilerplate migrations, test generation for pure functions, dependency bumps with a lockfile diff. What to keep human: auth, money, deletions, anything that touches production data.
Give the agent a small repository slice, not the monorepo. Context stuffing a million tokens of unrelated apps is how it "fixes" the wrong package.
Hiring metaphor that holds
You would not hand a junior engineer production credentials on day one. Do not hand SWE-2 your deploy token either. Separate identities. Separate secrets. A kill switch.
SWE-2 is a lever. Levers still need a hand that knows when to stop pressing.
Written by
Kirtesh
Founder
Kirtesh is a software engineer, indie hacker, and tech analyst writing on bootstrapped micro-SaaS, autonomous AI agents, cloud architectures, and the mechanics of building profitable software businesses.
