● THE COGNITION ANNALS · node 07 · uplink secure · channel: DOGBED · knap → k4o → textile → oliver → html

INDEX COGNITION DEVIN WINDSURF PRODUCTS SWE-1 SWE-1.5 SWE-1.6 SWE-1.7 SWE-2 TIMELINE LEDGER COMMITS

transmission 08 · model wing, hall iv

SWE-1.7: HARNESS-NATIVE

July 8, 2026. The crossover point: a model trained directly inside the Devin agent harness, on tasks long enough that a single rollout blows past the context window. SWE-1.7 was the lab’s first claim to frontier-level intelligence — at a fraction of frontier-level cost.

TRAINED WHERE IT WORKS

Every SWE model before this one was trained for a harness. SWE-1.7 was trained inside one. The Devin agent loop — plan, act, observe, retry — became the RL environment itself, which is a different claim than ‘evaluated in Devin’: the gradient went through the real thing.

THE TWO PROBLEMS LONG-HORIZON RL BRINGS

Context exhaustion — Devin tasks run long enough that rollouts exceed the raw context window; training had to teach the model to work beyond what it can literally see.
Reasoning bloat — RL on reasoning tasks tends to inflate response lengths (the DeepSeek R1 effect). SWE-1.7 was trained to elaborate only when the task is hard, keeping routine work cheap.

THE NUMBERS ON THE PLACARD

field	entry
released	july 8, 2026
claim	frontier-level intelligence at a fraction of the cost
training	RL directly in the Devin harness, long-horizon tasks
serving	cerebras at 1000 tokens/sec
variants	SWE-1.7 · SWE-1.7 Lightning (the cerebras speed tier)
contemporaries	kimi k2 · gpt-5.5 · glm · composer-1.5

THE HISTORICAL FOOTNOTE

SWE-1.7’s most durable contribution to the archive may be as a baseline. Two months later, the SWE-2 announcement measured itself against this model: 58% fewer turns, 81% lower cost, first code edit at step 18 instead of step 48. Every artifact in a museum eventually becomes the ruler a newer artifact is measured with.

From the start, we built Devin for completing asynchronous, long-running tasks. SWE-1.7 is trained directly in the Devin harness. — Cognition, July 2026

Final hall: SWE-2 — the model this lineage was rehearsing for.

— END OF TRANSMISSION · CTR-(ok)VERIFIED% —

INDEX COGNITION DEVIN WINDSURF PRODUCTS SWE-1 SWE-1.5 SWE-1.6 SWE-1.7 SWE-2 TIMELINE LEDGER COMMITS

+++ COGNITION ANNALS +++ DEVIN · THE AI SOFTWARE ENGINEER +++ SWE-1 → SWE-1.5 → SWE-1.6 → SWE-1.7 → SWE-2 +++ WINDSURF FOREVER +++ COMPILED BY DOGBED, NOT WRITTEN BY HAND +++ NO JAVASCRIPT WAS HARMED IN THE MAKING OF THIS ARCHIVE +++

rendered by dogbed — a document compiler, not a CMS. one binary, one template, thirteen documents.