Compyle

Compyle is a question driven AI coding agent built for complex, open ended work in real codebases, and its central bet is collaboration over full autonomy. Where many agents run off into long autonomous sessions and hand back large blocks of code the developer barely understands, Compyle deliberately keeps the human in the driver's seat. It asks targeted clarifying questions before and during coding, confirms decisions, and aims to produce code the developer actually understands and can build on, avoiding two thousand line black boxes and full rewrites. It is built on top of Claude Code and is backed by Y Combinator, though it is still an early stage product with a short track record.

The workflow has a deliberate shape. Every task starts with a research phase in which the agent scans the existing codebase using pattern matching, database schema checks, and file traversal across one or more repositories to learn the project's conventions. It then produces research and planning artifacts and works with the developer to agree a detailed plan, requiring explicit approval for architectural decisions before writing anything. During implementation, a system Compyle calls the Overwatcher continuously validates generated code against the agreed plan and the project's defined patterns, automatically flagging conflicts such as clashing database schemas and pausing to ask rather than pressing ahead on assumptions.

Compyle fits into an existing engineering workflow rather than replacing the editor. It connects directly to repositories, manages branches, generates pull requests, communicates bidirectionally with CI/CD pipelines, and has access to local files, builds, and tests, and it exposes an API. Teams can define their patterns and best practices once so the agent enforces them on every change, which is aimed at reducing technical debt and simplifying code review. It is model agnostic, running across GLM 4.6 and Anthropic's Claude Sonnet and Haiku models and orchestrating between them for different kinds of work.

Compyle launched with a free forever plan, with additional paid tiers that are not yet fully public as the product matures. It is best suited to software engineers and technical leads building substantial features, refactoring, or changing database schemas in mid to large codebases, and to teams moving from prototype to production who care about architectural consistency. The tradeoff is inherent to the design: because it asks questions and validates continuously, it requires active involvement and can feel like overhead on small, well defined, repetitive tasks where a more autonomous agent would simply finish the job.