Building agents in production
Shipping agent graphs without burning your margin budget
On this page
Autonomous agents are cheap to demo and expensive to keep honest. The moment you wire “just one more tool” without a contract, your margin graph looks like a crypto wallet after a bridge hack.
1. Budget the graph, not the model
Most cost spikes are not raw token spend — they are unbounded fan-out and orphaned retries. Before choosing an orchestration library, answer three numbers on paper:
- Max tool calls per user intent
- Max wall-clock per agent hop
- Max dollars per successful outcome (not per request)
If you cannot answer those, you are not ready to productionize — you are cosplaying infrastructure.
2. Evaluators are part of the DAG
Treat evaluators as first-class nodes, not weekend notebooks. They should run on the same artifacts as production (prompt versions, memory snapshots, tool traces) and block promotion when drift exceeds your policy.
3. Memory is a liability ledger
Persistent memory is not “helpful context” — it is debt. Every write needs an owner, TTL, and audit strategy. Otherwise you will spend Q3 hunting silent poisoned entries that only appear on long-horizon tasks.
4. Multilingual leadership note
I write in English on this site first because it is the lingua franca of the teams I advise day to day. When I cross-post externally, canonical URLs stay anchored here so Google (and recruiters parsing JSON résumés) always see one source of truth.
Want this delivered with smaller essays? Subscribe via the footer embed.