A convincing agent demo proves that a model can succeed on a prepared example. It does not prove that a workflow is dependable when requests are incomplete, policies conflict, tools fail, or a user's situation falls outside the examples. The evaluation ladder turns that gap into a sequence of evidence a venture can collect before it grants an agent more responsibility.
Demos hide the distribution
Production work includes ambiguity and variation. A useful evaluation set represents the requests that arrive, the edge cases that create real cost, and the policies that constrain the business. It should be held apart from the examples used to design the workflow. Otherwise the agent is being graded on the answers it has already seen.
A four-rung ladder
Each rung answers a different question. Skipping rungs produces false confidence.
Core idea: Autonomy is earned through observed operation, not inferred from model capability.
- Replay: can the workflow produce acceptable outputs against a curated, held-out set of past cases?
- Shadow: can it run beside the current process, without acting, and reveal where it agrees, misses, or needs context?
- Supervised: can an operator approve or correct its work within a useful response time and review burden?
- Constrained execution: can it act inside narrow policy limits with logging, rollback, and exception handling?
Measure the decision, not the prose
Fluent output can conceal a wrong routing decision, a missing policy condition, or an unsupported claim. Define the evaluation around the workflow's consequence. For a support triage agent, that may be correct routing, urgency, evidence shown, and timely escalation. For a research agent, it may be source quality, traceability, and whether a human can safely use the recommendation.
- Task quality: did the workflow reach the correct or acceptable outcome?
- Evidence quality: can a reviewer see the source, rule, or uncertainty behind the output?
- Operational quality: did it complete within the expected time, cost, and tool reliability limits?
- Escalation quality: did it stop and route the work when evidence or policy was insufficient?
Build the review loop into the launch
Create a representative case set
Collect normal, difficult, and policy-sensitive cases from the real workflow with a human-approved outcome.
- Record what makes a case hard
- Review data access and retention before use
Set thresholds before the test
Decide what counts as acceptable quality, what requires review, and what blocks release before results arrive.
- Separate reversible from irreversible actions
- Set a limit for unsafe or unsupported outcomes
Run in shadow
Compare agent proposals with the existing human process without changing customer-facing outcomes.
- Classify disagreements
- Turn repeated corrections into workflow changes
Widen only with evidence
Move a narrow action into supervised use, then constrained execution, only after the error and review pattern is understood.
- Keep rollback available
- Review the threshold at a fixed cadence
The ladder keeps an agentic venture honest. It makes performance visible before the organization builds dependency around it, and it gives founders a credible way to say both what the system can do today and what evidence it still needs.