erstan-agent-review

Erstan Agent review

Objective

Reconstruct what the Agent was configured to do, what the runtime actually did, and where the observed behavior diverged. Base every finding on an Agent definition, run record, trace event, persisted conversation item, tool result, or reproducible request.

Resources

• Read the Agent checklist when reviewing the graph, model, instructions, Skill policy, tool policy, or write boundary.
• Read the run checklist when diagnosing a run, trace, stream, persisted output, approval, retry, or tool effect.
• Read fix boundaries before assigning a remedy to the Agent, Skill, platform, or connector.

Workflow

1. Establish the requested scope: review, diagnosis, proposed change, or an authorized implementation.
2. Capture stable identifiers for the workspace, Agent, version, run, thread, and relevant tool calls without copying credentials into reports.
3. Read get_agent_builder_guide and the relevant detailed fields from get_agent_node_catalog, then inspect the Agent with get_agent, list_agent_tools, and list_agent_skills. Call get_agent_skill only for a workspace Skill ID. Never pass a system:<key> ref: system Skill packages are not readable through that tool. Record that evidence limitation instead of treating it as a missing workspace Skill. Do not infer current schemas from screenshots.
4. Run validate_agent when the connected app has Build agents, then apply the Agent checklist. Validation is non-mutating, but its current permission group is Build agents, which includes the required View agents dependency. Users manage that access in Settings > Connected apps.
5. For a run issue, read get_run and page get_run_trace to the latest terminal event, then apply the run checklist. Run evidence may be restricted to the exact credential or connection that launched it. A run_not_found from another connection is not evidence that the run never existed.
6. When the run came from the UI or another connection, review an authorized, downloaded local diagnostics directory: manifest first, every trace file second, persisted transcript third, and any HAR/SSE evidence last.
7. Compare live presentation with persisted conversation and trace evidence when the report involves disappearing, duplicated, truncated, or reordered content.
8. Classify each root cause using the fix-boundary guide. Separate primary cause, contributing conditions, and unrelated observations.
9. Recommend the smallest generic fix that preserves correct behavior for other Agents. Implement only when the request authorizes changes.
10. Validate the revised Agent, then exercise the smallest representative run. For writes, verify the business effect by stable identity before retrying.