Fully autonomous software engineering at scale
Descant is built by Autonomy to deliver consistent engineering quality across a sustained body of work. It handles planning, implementation, verification and independent review, and coordinates the dependencies and follow-ups across a project.
Engineering process
Descant investigates requirements, failure cases, architecture and integration seams. It reviews plans, verifies behavior and looks for ways to reduce unnecessary complexity. The effort is sized to the assignment.
Explicit contracts, type-safe interfaces and schema-validated data connect model judgment to software-enforced processes. Requirements, review findings, resolutions and evidence are tracked through implementation and verification. Claims about tests and simplification require evidence.
Coordination
Descant creates scoped follow-up issues, picks them up and coordinates them with existing work. Independent work can run in parallel; dependent work follows an explicit order. Related issues and learnings are maintained as the code changes. Issuegraph, the MA2H protocol and OH HAI support this coordination and communication.
Operating evidence
The research prototype shipped more than 20,000 pull requests before the production version was complete. Full transcripts are used to identify repeated errors, missed instructions, integration gaps, technical debt and unnecessary token spend, and to test improvements to the process.
An early Descant run scored 86.6% on SWE-bench Pro, compared with **80.3% for Mythos 5**, the leaderboard leader at the time. The underlying models used in that Descant run collectively scored 62% on their own. Our latest published result is 82.3% on DeepSWE v1.1, compared with 74.2% for Opus 5.5 and 79% for Muse Wonder Full, the leader on the LLM Stats leaderboard as of November 1, 2026. That is a 3.3-percentage-point lead on DeepSWE v1.1. That 3.3-point lead represents approximately 14.8 SWE-bench Pro points of time-equivalent progress, calculated from fitted leaderboard trends. Data and calculation
A benchmark does not establish performance across a sustained project. Evaluate the implementation, review history and completed work before committing to Descant.
Company and development
Autonomy was founded by Tim Layton, whose engineering career includes Shopify, Amazon and Boulevard. Tim leads product and engineering. Autonomy uses Descant to develop its own products.
As models improve, we intend to remove machinery that compensates for their limitations. Purpose, information, tools and coordination remain necessary. Our development effort can then focus more on expanding what customers can accomplish.
Talk about your engineering work
See the quality for yourself on your own codebase.