A software delegation contract is a proposed unit of analysis for coding work that is delegated to an AI coding agent: it covers the task being assigned, the authority the agent is given to modify the repository, the work package the agent returns, and the acceptance context in which that work is reviewed. It frames the agent as accepting an assigned task, modifying a repository under bounded authority and returning work for review.
Usage
The term comes from research on AI coding agents. According to Software Delegation Contracts: Measuring Reviewability in AI Coding-Agent Work, the concept was proposed in prior work without its effects being measured; that paper uses it to write explicit contracts for coding agents and compares them against a realistic issue-style prompt, both with and without a required evidence bundle.
When It Applies
The concept applies when a coding agent is assigned a bounded task and returns work for someone to accept or reject. Its evidence so far is one controlled pilot study of 64 agent runs on ten small TypeScript API tasks, reviewed by model-based reviewers. In that study, making the contract explicit did not improve correctness — every run passed the hidden acceptance checks — but improved reviewability: evidence sufficiency rose, reviewer ambiguity fell, and changed-file lists, known-limitations and residual-risk sections, and reviewer checklists appeared mostly or only when the contract demanded them. The cost was 13% more agent tokens and 38% more wall-clock time, larger for the weaker model tier.