XNews

Addy Osmani Outlines Self-Verifying Coding Agent Workflow for Managers

Addy Osmani Outlines Self-Verifying Coding Agent Workflow for Managers

Addy Osmani outlined a workflow for managing coding agents that prioritizes self-verification, requiring each task to include executable checks such as tests, builds, type checks and screenshot validation before an agent submits results with evidence. He advised supplying business context and success criteria that code does not capture, cleaning up outdated rules as models improve, and defining completion conditions, permissions and approvers before running unattended tasks, with humans retaining product, safety and release decisions.

Event version 1.

Key insights

Self-verification comes first

Each task should provide executable checks such as tests, builds, type checks and screenshot validation for interface changes. Agents should execute, check and fix failures before submitting results with evidence, and a fresh agent can independently review work when needed.

Context beyond the repository matters

Task instructions should explain why the work is done, business constraints, what not to do and what counts as success. Repeated corrections can become shared rules or skills, but rules should be cleaned up as models improve, with preferences written as guidance and mandatory safety requirements enforced.

Human attention limits parallelism

Unattended tasks require defined completion conditions, verification methods, permissions, run limits, deliverables and approvers. New processes should be piloted at small scale before expansion.

Humans retain product and release judgment

Passing tests do not make a feature worth building; product trade-offs, safety boundaries and release decisions remain with people. Effectiveness should be measured by delivery cycle time, rework and rollbacks rather than lines of agent-written code.

Source posts