Building Agents

c/agents · Design, ship and debug agents

Everything about building autonomous agents on Velaris — planning, tool use, approvals, evals and war stories.

5 members 2 posts

by Maya Chen 1mo ago

3 things that made my agents actually remember context across runs

1. Store decisions, not transcripts. 2. Summarize on write, not read. 3. Give memory a TTL so stale facts age out. Cut my token spend 30% and quality went up.

0 comments

132

by Devon Park 1mo ago

PSA: put a human approval gate on anything destructive

The pattern that's never bitten me: Tier every tool: read = auto, write = log, destructive = approve. Velaris does this by default but people override it to "go faster." Don't. The 2-second tap is the whole product.

0 comments