Failure PatternDecision layer

The wcagent Evidence Gap Trap: Why Agencies Fail With wcagent on Client Repos

Symptom: AI-generated diffs land in client pull requests with no attached test output, build log, or diagnostics, so the client's reviewer cannot verify the change and bounces it back. Root cause: wcagent does not bundle model access. The $4.99 monthly seat fee covers the extension only, and every developer still needs an eligible third-party AI service account, which agencies routinely forget to budget or pass through to the client.

By InnovaAI ResearchPublished Updated

AI Code Tools
How do you recognize it?
  • •AI-generated diffs land in client pull requests with no attached test output, build log, or diagnostics, so the client's reviewer cannot verify the change and bounces it back.
  • •The agency's evidence history is empty or partial for a sprint even though developers ran wcagent tasks daily, because approvals were granted without completion gating enabled.
  • •Client invoices show wcagent at $4.99 per seat per month while the real cost per developer is higher once each developer's separate ChatGPT, Claude, Gemini, Grok, or DeepSeek subscription is added.
  • •Developers report that wcagent sends irrelevant files or misses key symbols, producing edits that compile locally but fail in the client's CI pipeline.
  • •A client asks for an audit trail of AI-assisted work on their repository and the agency can only produce chat transcripts, not recorded diffs tied to passing tests.
Why does it happen?
  • •wcagent does not bundle model access. The $4.99 monthly seat fee covers the extension only, and every developer still needs an eligible third-party AI service account, which agencies routinely forget to budget or pass through to the client.
  • •Completion gating is a configuration choice. If the agency leaves developer approval settings permissive and does not require passing tests, successful builds, or recorded diffs before a task is marked done, wcagent will accept model confidence as the finish line.
  • •Repository context ranking depends on project instructions. Without an AGENTS.md file and a curated context selection, the context engine cannot rank the right files and symbols, so the model works from a thin or wrong slice of the client codebase.
  • •The extension runs in desktop VS Code only. Agencies that assume it works in browser-based editors, CI runners, or other IDEs will find no execution path, and local tool approvals cannot be enforced where the extension is not installed.
How do you fix it?
  • •Open wcagent settings in VS Code and turn on completion gating so a task cannot be marked complete without a passing test run, a successful build, or a recorded diff.
  • •Add an AGENTS.md file at the root of each client repository with project-level instructions, then re-run context selection so the context engine ranks the correct files and symbols.
  • •Audit the eligible provider connections for every seat and reconcile the agency's AI service account costs against the $4.99 per seat per month wcagent fee before quoting a client retainer.
  • •Review the evidence history panel for the last two weeks of client work and backfill any task that lacks test results, build output, or diagnostics before the next client review.
Open the wcagent coursewcagent decision recordBrowse the Academy

More on wcagent

More for AI Code Tools