TPM Agent Ecosystem

A hierarchical multi-agent system built on Claude Code that handles the full spectrum of TPM daily operations — from morning briefings to cross-team program health monitoring. Now in its third week of production use, with 22 agents generating 60+ automated reports.

Architecture

The system uses a hierarchical dispatch model: one Orchestrator routes requests to specialized sub-agents based on task type. Model tiering matches cognitive load to cost — Opus for cross-source synthesis, Sonnet for structured/mechanical tasks.

tpm-team-lead (Orchestrator)
├── Opus agents (complex synthesis)
│   ├── daily-briefing       → morning status across all programs
│   ├── eod-summary          → end-of-day digest with blockers
│   ├── program-monitor      → cross-team dependency health
│   ├── risk-radar           → flags at-risk initiatives proactively
│   └── ai-portfolio-review  → 688-line 8-phase portfolio analysis
└── Sonnet agents (structured tasks)
    ├── sprint-board          → Jira board scrape + velocity delta
    ├── roadmap-sync          → initiative status vs. last week
    ├── standup-notes         → formats raw notes into Jira comments
    ├── initiative-check      → ADF staleness scan + nudge drafts
    ├── feedback-triage       → Slack feedback × Jira defect matching
    ├── meeting-prep          → agenda + pre-read from calendar
    ├── post-meeting-followup → action items from meeting notes
    ├── launch-readiness      → feature completion % across categories
    ├── squad-page-auditor    → wiki hygiene (found 6 stale pages)
    ├── permissions-auditor   → tool access review + drift detection
    ├── security-auditor      → 8-phase credential + config review
    └── +5 more specialists

MCP & CLI Integrations

Every agent connects to real data sources via Model Context Protocol and CLI tools:

Impact (Week 3)

Key Design Patterns

1. Model Tiering — Opus for synthesis, Sonnet for mechanical. Saves ~60% with no quality loss on structured tasks.
2. Scoped Execution — Run single pipeline phases instead of full agents. A KTLO analysis costs $1-2 vs $5-7 for full portfolio review.
3. Human-in-the-Loop — Autonomous reads, reviewed writes. Agent drafts 5 Jira comments, you post 1.
4. Report-First Architecture — Every agent saves .md locally, uploads to Google Drive as native Doc.
5. Agents Auditing Agents — Permissions-auditor and security-auditor review the system's own config.

Evolution

Week 1 — Foundation

Built the first 11 agents. Struggled with MCP syntax, Jira field discovery, Glean date filter bugs. Established plan-mode → build → test → iterate pattern.

Week 2 — Integration & Intelligence

Expanded to 17 agents. Installed desk CLI (52% token reduction). Introduced model tiering. Built portfolio-level analysis agents that reason across multiple programs.

Week 3 — Maturity & Governance

Grew to 22 agents. First full operational days where the fleet ran the daily cycle end-to-end. Built governance layer (permissions-auditor, security-auditor). Cost-optimized to $35-50/week. The system stopped being a project and became infrastructure.

Build Log

Daily Journal (Week 1)

• Day 1 — Zero to 35 Skills
• Day 2 — Building My First Three AI Agents
• Day 3 — When Sonnet Isn't Enough
• Day 4 — The Orchestrator Pattern
• Day 5 — The Day Everything Broke
• Day 6 — Scaling to Full Program Coverage
• Day 7 — Seeing What Your AI Actually Costs
• Day 8 — Cutting Costs by 40%

Weekly Summaries

• Week 1 — Building the Foundation
• Week 2 — TPM AI Operating System
• Week 3 — When Agents Start Auditing Themselves (coming soon)

Related

• The Orchestrator Pattern
• From Bash Scripts to Autonomous Agents
• Prompt Engineering for Production Agents
• MCP: The Protocol That Makes AI Agents Useful
• Replacing Glean with Direct Google Workspace Access