Working with AI Agents

Claude Code does most of the typing — but you are still the engineer. You decide what to build, you verify what the agent produced, and you sign off on the change going live. This article covers the habits that make the difference between "vibe coding that ships" and "vibe coding that breaks things".

The golden rule

Never merge code you don't understand at least at a high level.

You don't need to grasp every line. You do need to know what the change is supposed to do, where it lives, and roughly how to tell if it's working.

Giving your agent access to production observability

The Synapse MCP exposes Corti's Mimir (metrics) and Loki (logs) to your AI harness, so the agent can check whether a deployment is healthy, hunt error logs, or pull latency percentiles without you leaving the editor. Set it up in OpenCode in a few lines — see Configuring the Synapse MCP, and paste the companion agent skill so your harness knows the label conventions and query patterns.

Giving good instructions

A few habits that consistently produce better results:

Be specific about the goal, not the implementation. "Add a button that exports the table to CSV" beats "use the xlsx library and add a click handler".
Mention the context — which file, which page, which user.
Ask for small steps. One change, one PR.

Reviewing what the agent produces

Before committing, spend two minutes:

Read the diff in your editor or the PR view.
Ask Claude: "Explain what this change does in plain English."
Try the feature yourself locally if you can.

If something looks off, say so. Agents take corrections well.

Common pitfalls

Patterns to watch out for:

Scope creep — the agent "helpfully" refactors things you didn't ask about. Ask it to revert.
Made-up APIs — calls to functions that don't exist. CI usually catches these, but you might too.
Silent failures — code that runs but does nothing useful. Always test the outcome, not just that the code runs.

When to ask a human

Ask a colleague (not just the agent) when:

A reviewer's comment doesn't make sense to you.
CI fails in a way Claude can't fix after two tries.
The change touches something critical (auth, payments, data deletion).

The loop

graph LR
  P[Prompt] --> C[Claude codes]
  C --> R[You review]
  R --> P

Keep the loop tight. Many small prompts beat one giant one. Many small PRs beat one giant one. That's the rhythm — and the rest of this series is just the scaffolding that makes the rhythm safe.