Building good Agents
- Write a clear, testable AOP. It’s the highest-leverage thing you control. See Writing Effective AOPs.
- Match the model to the job. Harder reasoning and longer procedures benefit from a stronger model; simple, high-volume tasks don’t need one. See Model Selection.
- Decide one Agent or several. When a single AOP is juggling unrelated jobs, split it. See Multi-Agent Process Decomposition.
- Pick how Runs start deliberately. Schedules, triggers, Queues, and manual Starts each fit a different shape of work. See Choosing How Your Agent Starts.
- Design Queues so Cases always close. Every claimed Case should end completed, failed, postponed, or handed over. See Designing Work Around Queues and Cases.
- Give the Agent the right tools. Pick the Connection pattern that fits the system you’re integrating. See Choosing the Right Connection Pattern.
Running reliably
- Test in a revision, then promote. Validate changes safely before they reach production. See Test Safely Before Going Live and Promote to Production.
- Watch the first week. A daily review cadence catches problems while they’re cheap to fix. See Monitor the First Week.
- Plan for retries and failures. Decide what should happen when a step fails or input is malformed. See Retries, Failures, and Skipped Steps.
- Use Evaluations to find weak spots. Duvo scores Runs against criteria derived from the AOP. See the Runs List.
Staying safe
- Gate high-risk actions. Put caps, thresholds, and hard limits in the AOP for anything financial, irreversible, or outbound. See Guardrails for High-Risk Automations.
- Add human approval where it matters. Define exactly when the Agent should pause for a person. See Human-in-the-Loop and Designing Human-in-the-Loop Workflows.
- Share Logins, don’t keep them personal. A Run can use a Login that’s shared with the team (or owned by the person the Run runs as); a personal Login saved by someone else won’t be available to it. If a sign-in “isn’t loading,” check that the Login is shared. See Connections and Logins That Don’t Break.
Scaling
- Let Agents remember and learn. Persisted context and feedback make recurring work better over time. See Agent Memory and the Learning Feature.
We’re actively expanding this page. If a tip keeps coming up in support or in your own building, it belongs here.