Agent Task Audit Plugin
Make multi-agent work reviewable
When the strongest model coordinates and smaller workers execute modules, use hooks plus self-reports to record who did what, which tools were used, and which token counts are exact or estimated.
6
Demo Events
3
Worker Reports
1,238
Exact Tokens
Bilingual
Owner Report
What You Will Learn
- Explain why Codex/Claude multi-agent work needs auditability
- Separate automatic hook logs, subagent self-reports, and runtime visibility limits
- Use Agent Task Audit to record tool use, module ownership, and token visibility
- Turn raw logs into bilingual owner-readable reports
- Design a subagent self-report protocol for role, module, model, tools, files, blockers, and tokens
- Run the local demo viewer and import NDJSON audit logs
Course Materials
Submit to downloadSubmit a lightweight form to download the PDF, lecture notes, research notes, and demo screenshot.
To share materials with interested learners, leave your basic info and topics first. The file opens automatically after submission.
Course Flow
From opaque delegation to a working audit plugin and owner report.
Why Audit Agents
Goal delegation, coordinator models, workers, and owner opacity
Generality Boundary
Shared schema vs Codex/Claude-specific integration
Plugin Architecture
Hooks, self-reports, owner reports, and viewer
Self-Report Protocol
Role, module, model, tools, files, completed work, blockers, tokens
Demo
Import NDJSON and inspect owner summary and cost visibility
Productization
Claude hooks, cross-machine install, HTML/PDF reports
Record facts automatically and responsibility explicitly
Agent Task Audit is not a hidden-routing decoder. It is a ledger and handoff sheet for multi-agent work.
Standard Audit Loop
- Main agent receives a goal and declares audit requirements
- Hooks record prompts, tool use, permissions, and stop events
- Subagents execute module work and write self-reports
- Reports separate exact, estimated, and unknown fields
- Owners inspect summary, work allocation, cost visibility, and risks
- Technical teams inspect the raw timeline only when needed
3-Layer System
Hook Events
Automatic facts
Capture lifecycle events, tool calls, approval requests, and visible usage fields.
Self Report
Worker accountability
Each worker reports role, module, tools, files, completed work, blockers, and tokens.
Owner Report
Readable accountability
Explain who did what, where cost is visible, and what remains opaque in English and Chinese.
Owner Report Checklist
Every report should answer: 1. Who owned which modules? 2. Which model information was visible? 3. Which token counts were exact? 4. Which token counts were estimated or self-reported? 5. Which tools, commands, and files were used? 6. Were there failures, blockers, or unknowns?
Suggested Teaching Flow
| Time | Duration | Module | Mode |
|---|---|---|---|
| 00:00–10:00 | 10 min | The multi-agent opacity problem | Explain |
| 10:00–20:00 | 10 min | Generality and visibility limits | Clarify |
| 20:00–35:00 | 15 min | Plugin architecture and self-report protocol | Breakdown |
| 35:00–50:00 | 15 min | Demo viewer | Practice |
| 50:00–60:00 | 10 min | Productization roadmap | Discuss |
Who This Is For
For people using Codex/Claude in real delivery, development, research, and operations workflows.
Good Fit ✓
- AI automation consultants
- Technical leads and project owners
- Heavy Codex / Claude users
- Teams that need multi-agent accountability
Not a Fit ✗
- You expect hidden model routing to be decoded magically
- You do not want workers to self-report
- You only want raw JSON and do not need owner reports
FAQ
Is this plugin general?
The schema, self-report protocol, and owner report are general. Codex plugin installation and hook paths are platform and machine specific. Claude needs Claude Code hooks or a wrapper to write the same log format.
Can it see every true token count?
Only token counts exposed by the runtime are exact. Everything else is labeled estimated or unknown.
Why require subagent self-reports?
The main agent may not see every internal detail. Self-reports bring module ownership, deliverables, risks, and estimated cost back to the owner.
Want auditable AI delivery workflows?
We can help turn Codex/Claude multi-agent workflows into recorded, reviewable, owner-readable audit loops.