claude-agent-team.
A Claude Code setup for running your life with an agent team: an orchestrator, specialist personas for everyday domains, and the memory conventions that keep the team useful across months of sessions.
#What it is
A template you copy into ~/.claude: orchestrator rules, four starter personas (finance analyst, concierge, fitness coach, research analyst), and the memory files that hold the routing table and each agent's state. Most public Claude Code setups are built for writing software. This one is built for everything else: tax deadlines, contract cancellations, travel plans, training programs, the recurring life admin that benefits from an assistant with persistent memory. It is extracted from a setup in daily personal use.
finance-analyst
Personal finance: budgeting, taxes, investing basics, money paperwork
triggersbudget, tax, invest, ETF, broker, savings, debt, loan, subscription, net worth
personaThinks like a fee-only financial planner with an accounting background: conservative by default, allergic to products that mainly pay the person selling them.
hands offBooking travel, canceling contracts, appointment logistics: concierge.
> Where do I even start with my tax return this year?
Routed to finance-analyst.
template/agents/finance-analyst.md
concierge
Logistics and bureaucracy: travel, bookings, contracts, letters, deadlines
triggerstrip, flight, hotel, appointment, contract, cancel, insurance, complaint, refund, deadline
personaWorks like a senior executive assistant: takes annoying operational tasks off the user's plate end to end.
hands offJudging whether something is financially sensible (a switch, a big purchase): finance-analyst.
> The gym raised its prices again. I want out.
Routed to concierge.
The deadline is now tracked in the concierge backlog.
template/agents/concierge.md · docs/walkthrough.md
fitness-coach
Training programs, progression tracking, nutrition basics for training goals
triggersworkout, gym, training plan, exercise, running, cardio, protein, muscle, recovery
personaThinks like a certified personal trainer with a sports science background: skeptical of trends, loyal to the basics with evidence behind them (progressive overload, enough protein, sleep).
hands offCanceling or negotiating the gym contract itself: concierge.
> Three gym days a week. Build me a training plan for muscle.
Routed to fitness-coach.
template/agents/fitness-coach.md
research-analyst
Structured research: comparisons, fact checks, summaries, decision tables
triggerscompare, research, best option, reviews, verify, fact check, summarize, alternatives
personaWorks like a reference librarian with an analyst's habits: goes to primary sources, treats marketing copy as a claim to check rather than an answer.
hands offExecuting the winner (booking, ordering, canceling the old provider): concierge.
> Time for a new mattress. Compare the options and check the reviews.
Routed to research-analyst.
template/agents/research-analyst.md
agents/_template.md
Copy this to create your own agent
1Research what a real expert in the field knows (skills, vocabulary, failure modes).
2Create agents/<slug>.md from agents/_template.md plus an empty agents/<slug>-references/backlog.md.
3Add the roster row here, in the same commit or session.
totalThree files, roughly 30 minutes.
template/memory/team_roster.md · README.md
#The itch
A single generic assistant forgets. Every context clear costs the running state of your life admin, and every new session starts by guessing facts about you. A generalist also glosses over the domain details that change answers; the concierge persona here knows that a price increase can open an extraordinary cancellation route. The setup exists so that state, facts, and domain knowledge each live in a file that survives the session.
#How it works
Three mechanics carry the whole setup:
- Routing. The orchestrator checks the team roster and either answers directly or delegates to the persona whose trigger topics match.
- State of play. Every agent keeps a backlog file with its running work, next step, and deadlines, updated unprompted after each session. Context clears stop costing anything, because the state lives in a file.
- Hiring. When a recurring topic has no owner, the orchestrator researches the domain, writes a new agent file from the template, and adds the roster row, in one session. The team grows with your life instead of being designed up front.
trivial or general questions
answers itself
exact match on role or trigger topics
delegate
to the persona whose trigger topics match
no match and the topic will recur
hire
a new agent file from the template, plus the roster row
Agents do not call each other; every handoff runs through the main session.
The personas are starting points; replace them freely. The durable part is the convention set: reference folders that keep agent files small, state-of-play backlogs that survive context clears, a routing roster, and a decision rule for where knowledge gets saved. Those took months of daily use to settle, and each one is documented with the failure it prevents.
#In real use
The demo session in the README is labeled as what it is: a reconstructed example, not a pasted transcript. The other limits, from the repo:
- Built and tested with Claude Code on Windows. The layout is platform-independent but has had less testing elsewhere, and the statusline script is tested on Windows (Git Bash) only.
- Personas are deliberately generic. They get useful after you fill
memory/user_profile.mdand let the backlogs fill up. - Anything with real-world consequences (payments, cancellations, filings) stays with you by design. Agents draft and analyze.
- The conventions assume you actually clear context between topics. One endless session leaves the backlog files lagging behind reality.
Two releases so far: 0.1.0 (2026-07-17) with the full template, 0.2.0 (2026-07-29) with fixes that came out of first-time installs, like the quickstart failing when ~/.claude did not exist yet.
#Install
$ git clone --depth=1 https://github.com/belschak/claude-agent-team.git
$ mkdir -p ~/.claude
$ cp -rn claude-agent-team/template/* ~/.claude/
The copy skips files you already have, so an existing ~/.claude/CLAUDE.md stays untouched. Worth knowing before you run it: this is a global install, and the orchestrator rules load into every session in every directory, software projects included. The README explains how to confine the setup to one directory instead.