Every framework can call a tool. Almost none of them can tell you what happens when the tool returns garbage, the model hallucinates an argument, or a task takes forty minutes and three retries. Demos do not survive contact with production — agents that finish jobs do. Here is what we learned building Agent Runtime for customers who measure success in completed tasks, not chat turns.
Pattern one: verify, then retry
The cheapest reliability win is a verifier between the model and the world. Every tool call gets a schema check, every state change gets a post-condition, and failures return to the model with the reason attached instead of a generic exception. In our telemetry, verifiable retries fix the majority of first-attempt failures without a human in the loop.
An agent that cannot tell success from failure is a random number generator with better marketing.
Pattern two: structured tool contracts
Free-form arguments are where agents go to die. We ship tools with strict JSON schemas, explicit error envelopes and idempotency keys. When a model proposes a call that does not validate, the runtime rejects it before execution and feeds the schema error back — the model corrects itself on the next attempt in the large majority of cases.
- Idempotency — a retried call must not double-charge a customer or re-send an email.
- Timeouts with budgets — every tool declares a deadline; the agent plans inside it.
- Dry-run mode — the whole plan is previewable before anything executes.
Pattern three: human checkpoints where they count
Autonomy is not binary. We let teams mark specific steps as checkpoints — anything that spends money, sends external communication or deletes data. The agent runs to the checkpoint, presents its evidence, and waits. Most workflows end up with two or three checkpoints, not thirty.
Evals are the product
You cannot improve what you cannot measure. Every agent we ship comes with a task suite: fifty to a few hundred real workflows with graded outcomes, run on every prompt or tool change. When the pass rate drops, the build stops. That discipline, more than any model choice, separates agents that demo well from agents that work.
What this means for your rollout
Start with one workflow that has a clear definition of done and a human fallback. Instrument it. Add checkpoints as trust builds. The teams that succeed treat agent deployment like hiring a new employee — onboarding, supervision, then autonomy — not like installing software.