Tips & Best Practices

A few practices that make atomic-agent faster and more reliable in daily use.

Keep the prefix stable

The runtime reuses the KV-cache only while the prompt prefix is byte-stable. Avoid churning the persona, rules, or tool set mid-session — let the stable prefix do its job and you’ll feel the speedup.

Let independent reads run in parallel

When you ask for several things that don’t depend on each other (read three files, check two pages), the agent can batch them after a single inference — up to 4 at once. Phrase requests so independent lookups are obvious.

Use approval gates deliberately

Dangerous operations ask first. Keep approvals on for anything that writes files or runs shell commands until you trust a given workflow, then streamline.

Read the traces

Every step is logged to local NDJSON. When something goes sideways, replay the trace to see exactly what the model saw and did — it’s the fastest way to debug an agent run.

Capture repeatable work as a skill

If you find yourself walking the agent through the same procedure twice, write it up as a skill so it loads on demand next time.