Thomas De Vos 👍

After a decade building AI systems for financial institutions, I wrote the book I wish I had when I started with agents. "Claude Code: Building Production Agents That Actually Scale" is live on Leanpub. leanpub.com/claude-code-buil… #ClaudeCode #AIAgents #Anthropic #AIEngineering

A clean diff is not a review. Before I approve, I ask for objective, permission boundary, tool trace, tests, retries and a rollback path. Put review and rollback checks in place before approval with my Claude Code book's ninety-day readiness checklist: thomasdevos.com/go/claude-co…

A timeout is not a failure. It only means the client stopped waiting. My fix: one operation ID per task, checked before any retry. That pattern and 31 chapters on Claude Code agents in production are in my book: leanpub.com/claude-code-buil…

Claude Code can finish a repair while its temporary production role still works. Close the run only when every lease is resolved and a canary with the retired principal is denied. More production controls: leanpub.com/claude-code-buil… #ClaudeCode #AgenticCoding

Do not let Claude Code explain its patch to the evaluator first. That can turn review into agreement. Use two passes: patch and raw evidence first, author explanation second. thomasdevos.com/posts/2026-0… #ClaudeCode #AgenticCoding

If Claude Code writes the patch and the eval, a green score may only prove both share the same mistake. Freeze the eval before the run. Keep hidden denial cases outside the agent's context, version the evaluator, and record which suite judged the patch. #ClaudeCode #AIEvals

File scope is a weak boundary for Claude Code. A CI change can deploy an image, rotate a secret or run a migration. Approve the effect, target and rollback path, not just the pathname. More production agent checks: leanpub.com/claude-code-buil… #ClaudeCode #AgenticCoding

A clean Claude Code replay can still lie when telemetry drops a production tool call. This trace receipt checks sequence gaps, terminal outcomes, observed effects and sink acks. Any mismatch blocks replay and review. thomasdevos.com/posts/2026-0… #ClaudeCode #AIAgents

Treat merge conflict resolution as a new agent task, not clerical cleanup. A conflict can combine two individually safe changes into a new behavior. Re-state the file scope, rerun boundary tests, and issue a fresh review packet before merge. #ClaudeCode #AgenticCoding

A production-agent handoff should say this plainly: EXPECTED 42 COLLECTED 41 GAPS 1 REPLAY BLOCKED The trace is incomplete, so the review stops. I cover this and other production controls in my Claude Code book: leanpub.com/claude-code-buil… #ClaudeCode #AgenticCoding

Claude Code can lose a safety decision when its context is compacted. If the denial lived only in the transcript, a summary can forget it. I keep control state outside the model and run a restore gate before tools come back: thomasdevos.com/posts/2026-0… #ClaudeCode #AgenticCoding

An agent can finish the code and still fail the run by spending its budget on implementation. Reserve capacity for tests, diff review, rollback notes and the handoff. If that reserve is touched, stop adding code. A production run needs an exit budget. #ClaudeCode #AgenticCoding

When a Claude Code change misbehaves, cap repair attempts before the run: one diagnosis pass, one bounded fix, then rollback. Otherwise, recovery becomes an agent exploring live systems under pressure. Put the stop condition in the approval record. #ClaudeCode #AgenticCoding

A timed-out MCP write might already have committed. If Claude Code retries with a new operation ID, one missing response can become two production effects. I wrote up the operation journal and retry gate I use to prevent that: thomasdevos.com/posts/2026-0… #ClaudeCode #MCP

A Claude Code review packet can name the commit, diff and passing tests, then stop one step too early. Add the artifact digest deployed and the digest observed at runtime. If reviewed, deployed and observed identities split, the release is not verified. #ClaudeCode #DevOps

Before Claude Code writes through MCP, make the tool require the version it read. If the record changed, reject the write and force a fresh plan. That compare-and-set check closes the gap between approved state and state at execution. #ClaudeCode #MCP

Claude Code can read three correct production facts and still plan against a state that never existed. I wrote up a coherent read-set receipt that catches mixed snapshots before an MCP action: thomasdevos.com/posts/2026-0… #ClaudeCode #MCP

A Claude Code eval passed because the mock MCP tool returned the expected JSON. In production, the same response also queued a background write. Test the observed effect, not only the response body. A convincing mock can hide the exact failure you need to catch. #ClaudeCode #MCP