Building a TPM Agent Ecosystem with Claude Code

It started with a simple question: What if my morning routine could run itself?

As a TPM managing multiple workstreams at my company, I was spending 2+ hours every morning just getting caught up — reading Slack threads, checking Jira tickets, pulling sprint data, and prepping for meetings.

The Problem

Every morning looked the same:

• Open 10+ Slack channels, skim for overnight activity
• Pull up Jira, check sprint board status across 3 workstreams
• Cross-reference roadmap initiatives with sprint goals
• Prep talking points for 4-6 daily meetings
• Draft standup updates for team channels

This was 2 hours of pure context-switching before I could do any actual TPM work.

The Architecture

I built a hierarchical dispatch system where a team lead orchestrator routes compound requests to specialist agents:

tpm-team-lead (orchestrator)
├── tpm-daily-assistant (Opus) — calendar-driven morning briefing
├── tpm-eod-summary (Opus) — end-of-day synthesis
├── dct-program-monitor (Opus) — cross-team program tracking
├── ai-pillar-monitor (Opus) — AI pillar status across Slack/Jira/Confluence
├── daily-update-publisher (Sonnet) — Slack standup text from Jira
├── sprint-board-publisher (Sonnet) — sprint data → Confluence
├── roadmap-publisher (Sonnet) — Jira ROADMAP → Confluence
├── standup-note-to-jira (Sonnet) — Gemini notes → Jira comments
└── ... 9 more specialists

Tiered Model Assignment

Not all agents need the same intelligence:

• Opus: Complex synthesis agents that cross-reference multiple data sources
• Sonnet: Structured/mechanical agents that follow templates

This saves ~60% on token costs while maintaining quality where it matters.

Key Agents

tpm-daily-assistant

Reads my Google Calendar, enriches each meeting with Jira/Confluence/Slack context, pulls inbox highlights from labeled emails, and generates an actionable daily agenda with meeting prep notes.

dct-program-monitor

Scans 10 Slack channels, pulls Jira ROADMAP initiatives, reads Confluence tracking pages, and generates a cross-referenced digest with risk flags. Week-over-week context from cross-team meeting notes.

sprint-board-publisher

Pulls all tickets from the active sprint, groups by status, calculates velocity metrics, and publishes a formatted report to Confluence.

Results

Lessons Learned

1. Prompt quality > model quality — When agent results were suboptimal, the fix was always the prompt, never upgrading the model.
2. DB-first with static fallback — Every content function tries the database first, falls back to constants. Zero downtime during DB migrations.
3. Agent instruction adherence — Sonnet agents need strong, explicit language. "ALWAYS use desk first" works. "Prefer desk" doesn't.
4. Compound workflows — "Get me caught up" dispatches 3 agents in parallel. Single requests go directly to specialists.