Agent Observability · Demo verified

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.

Download materials

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 download

Submit 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.

Now
Week 01

Why Audit Agents

Goal delegation, coordinator models, workers, and owner opacity

Week 02

Generality Boundary

Shared schema vs Codex/Claude-specific integration

Week 03

Plugin Architecture

Hooks, self-reports, owner reports, and viewer

Week 04

Self-Report Protocol

Role, module, model, tools, files, completed work, blockers, tokens

Week 05

Demo

Import NDJSON and inspect owner summary and cost visibility

Week 06

Productization

Claude hooks, cross-machine install, HTML/PDF reports

Core Method

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.

Useful as tooling and operating-process training
Recommended 45–60 minutes
PDF + Demo
Report + Markdown + Screenshot

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

TimeDurationModuleMode
00:00–10:0010 minThe multi-agent opacity problemExplain
10:00–20:0010 minGenerality and visibility limitsClarify
20:00–35:0015 minPlugin architecture and self-report protocolBreakdown
35:00–50:0015 minDemo viewerPractice
50:00–60:0010 minProductization roadmapDiscuss

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.

Download materials