Problem context
Your team uses one or more coding agents heavily. Refactors and new features often come from agent output.
You have seen analytics calls disappear after agent-written PRs even when the rest of the diff looks clean.
The GitHub Action catches it, but at that point the human reviewer is already cleaning up after the agent.
What breaks if this is not solved
- Reviewer time. Catching drift after the agent commits means the human is now the corrective signal.
- Agent quality perception. Repeated post-hoc fixes make agents look worse than they are.
- Velocity. Each agent PR with instrumentation drift turns into a multi-round review.
When this playbook applies
- Your team has at least one agent that supports MCP (Cursor, Claude Code, Codex, Devin).
- Developers are willing to add a server to their agent's MCP config.
- You already have the GitHub Action in place as a fallback.
System approach
Two layers of validation, same engine. The MCP catches drift as the agent writes; the Action catches anything that bypassed the MCP.
Surface findings inside the agent's own loop. The agent should see and react before commit.
Keep the friction low: the install should be a one-line config change per developer.
Execution steps
- Install the MCP server: `npm install -g @skene/mcp`.
- For Cursor: add `{ "mcpServers": { "skene": { "command": "skene-mcp" } } }` to `~/.cursor/mcp.json`.
- For Claude Code: same structure in the Claude Code MCP config.
- Restart the agent. Confirm Skene appears in the available tools.
- Test: ask the agent to refactor a function that contains an analytics call. Verify Skene's tool gets invoked and reports drift if the call disappears.
- Share the install snippet in the team Slack or onboarding doc so every dev gets it.
- Keep the GitHub Action turned on as a fallback - some PRs will come from devs who have not installed the MCP, or from external contributors.
Metrics to watch
Agent PRs with no Skene findings
Should rise as agents learn from MCP feedback in-loop.
Time between commit and merge for agent PRs
Trend down as fewer review cycles are needed.
Failure modes
- Treating MCP as a replacement for the GitHub Action. Use both.
- Letting individual devs install the MCP without team-level visibility. The agent learns; the team should know what it is learning.
- Not periodically auditing what the agent has been accepting as 'no drift'. The agent's judgment is good but not infallible.
Related concepts
Glossary
Adjacent playbooks
