Evidence-Layered Evaluation for Interactive Agents

Interactive agents can reach the wrong outcome for many different reasons: a bad plan, an incorrect click, a transient site response, or an evaluator that cannot observe the relevant state. A single success/failure score hides these causes and makes regressions difficult to reproduce, especially when tasks run on live websites or desktop applications.

Capture evaluation evidence in layers, while keeping the task outcome as the primary decision signal. A run records (1) the final-state or request-level assertion, (2) the ordered actions taken, (3) screenshots or other visual state, (4) a session replay when timing matters, (5) network requests when the browser state is not sufficient, and (6) the agent messages that led to each action. A verifier can then report the outcome and link it to the smallest evidence layer that explains the result.

A practical pipeline is:

task + policy -> isolated run -> action/visual/network/message capture
                              |
                              v
                         outcome verifier
                              |
                     score + evidence bundle

The layers should be timestamped and correlated by run ID. Collectors may be disabled for privacy-sensitive tasks, but the evaluator should state which layers were available rather than silently treating missing evidence as a successful run.

• Evidence Grade: medium
• Most Valuable Findings: Separate evidence layers support diagnosis of planning, interaction, and environment failures; request-level assertions can verify state transitions that are not visually exposed; replayable bundles make human adjudication and regression analysis practical.
• Unverified / Unclear: The relative value of each layer depends on the task and site; no universal layer ordering or storage budget has been established.

1. Define a task-level success assertion before running the agent, such as a state change or intercepted request.
2. Run the agent in an isolated browser or desktop environment and attach a stable run identifier to every event.
3. Record actions and agent messages by default; add screenshots, video, and network capture when they answer a known observability gap.
4. Evaluate the assertion first, then use the evidence bundle to classify failures and adjudicate ambiguous outcomes.
5. Retain a compact manifest with timestamps, harness/model versions, task version, and enabled evidence layers so another evaluator can reproduce the comparison.

• Pros: More actionable failure analysis; easier regression triage; supports both automated verification and human review; makes claims about live-task performance auditable.
• Cons/Considerations: Storage and instrumentation overhead increase with each layer; network and message logs can contain sensitive data; clock drift and incomplete capture can complicate correlation; richer evidence does not guarantee a correct evaluator.

• ClawBench — an example implementation of layered evidence capture for live-web agent evaluation (contributor-maintained; cited as an implementation, not as a normative standard).
• BrowserGym — web-agent environments and evaluation tooling.
• WebArena — realistic web environments for autonomous-agent evaluation.