AGENTS
The bar for agents moves from demo to verified execution
A convincing agent run needs more than a fluent transcript. It needs a predeclared definition of done, evidence from the tools and systems it changed, and a record a human can audit.
What is verified
The OpenAI Agents SDK records model generations, tool calls, handoffs, guardrails, and custom events as traces and spans. Comparable observability is appearing across agent frameworks because a multi-step run cannot be understood from its final sentence alone.
A trace is evidence of execution. It is not proof that the execution was correct. A tool call can succeed while changing the wrong record; a test can pass while the requested behaviour remains incomplete; an agent can produce an excellent summary of work it did not finish.
The higher bar
Verified execution joins three things that demos usually separate:
- A specification written before the run. The task, constraints, allowed systems, and done signal are fixed before the agent begins.
- Evidence from the target system. Tests, diffs, API responses, rendered pages, or state queries confirm the result independently of the agent’s narrative.
- An auditable record. Tool use, failures, retries, human interventions, and unresolved risks remain available for review.
The record also needs a privacy boundary. Agent traces may contain prompts, model outputs, and tool inputs or outputs; the SDK documentation explicitly notes that these fields can be sensitive and provides controls to exclude them.
What we are watching
The useful comparison between agent systems will be completion rate under the same specification, not how autonomous a demo appears. Our Agent Field Test is being designed around that standard.