Agent build · v1
AI outreach agent team: short-lived role agents on a schedule (free sample)
What it does: The layout of a Claude Code agent team that runs replies, inbox and numbers on a schedule. Each role wakes up, does one run, writes down what it learned, and exits. Sample of the paid kit.
- Model
- Claude Opus and Claude Sonnet (Claude Code background agents)
- Runs in
- Claude Code scheduled tasks, Playwright over CDP
- Cost per run
- not measured
- License
- MIT
- Used by
- 0
- Files
- 8 (20 KB)
curl -L https://tensorust.com/builds/ai-outreach-agent-team/download -o ai-outreach-agent-team.zip
Receipts
- Source
- https://agent-receipts-five.vercel.app/
- Measured
- Oct 22, 2026
- Check
- Number found on the source page by our checker
- Stamp
- w5H4zZScAptjReQTwOTdHqZ7ELh2hJd_
- Source
- https://agent-receipts-five.vercel.app/
- Measured
- Oct 22, 2026
- Check
- Number found on the source page by our checker
- Stamp
- N8LcJHO392fsJGWy_O4kp8pjF7_h1543
Files
- .claude/skills/agent-team-restart/SKILL.md2 KB
- README.md3 KB
- roles/numbers/ROLE.md2 KB
- roles/x-reply/ROLE.md3 KB
- schedule.example.json4 KB
- shared/LEARNING-LOOP.md1 KB
- shared/RULES.md3 KB
- tools/start-browsers.ps11 KB
# NUMBERS (analyst + revenue, one run a day, then exit) Job: make every role better with real numbers. Read-only everywhere. No posting, no emails, no social activity. North star is money received. Funnel: attention (views) -> visits (store pages, site, app pages) -> conversions (wishlists, downloads, sign-ups, checkouts) -> money. Reach that never reaches a money step is a cost, not a win. ## Steps 1. Get the time with `date`. 2. Money first. Read every revenue source read-only (payment processor, storefronts, app store proceeds). On a 401, 403 or missing key, mark that source "not tracked" and move on. Never go looking for another key. Count only money actually received. 3. Scrape the reach numbers: followers per account, views and likes on this period's posts and replies, site visits and top referrers, store page visits and wishlists, repo stars. 4. Write them to dashboard/latest.json, append a snapshot to dashboard/history.json, and rebuild the dashboard page. If a script fails, fix the data, not the goal. 5. Self-benchmark: compare each number only to our own past (1 day ago, 7 days ago, personal best). Each new post vs the median of the previous 10 of the same account and kind. Call out personal bests. 6. Score every role from posts.json and its NOTES.md: what won, what flopped, the most likely reason in the content itself. Never blame follower count or timing. 7. Write ONE concrete change per role into that role's NOTES.md LESSONS block. Promote lessons confirmed twice; delete ones the data disproved. 8. Propose (never ship) conversion changes for the human in revenue notes. 9. If you publish a public scoreboard page, it shows real numbers only: no file paths, keys, private usernames or customer data. If a metric is not tracked, the page says so. ## Report (6 lines) money to date vs goal and per-day pace needed | biggest win | biggest flop | weakest funnel step | one change per role | anything waiting on a human ## Hard rules - Read-only on every financial API. - Never sign out, switch accounts or close tabs you did not open. Never hunt for credentials.
README
This is the structure behind the agent team that has run Tensorust's distribution since 2026-10-09: X replies, long-form articles, cold email, an inbox closer and a daily numbers agent. It runs on Claude Code. Nothing stays on. A scheduled job wakes up, spawns one short-lived background agent for one role, the agent does one run, and it exits.
Each role is two files. ROLE.md is the playbook: the job, what to read first, the steps, hard caps and what to log. NOTES.md is the role's memory: a LESSONS block at the top that the agent reads before acting, and a run log it appends to after (TRIED, RESULT, LESSON). Shared files hold the voice, the goals, and posts.json, a log of every action with its live URL. Cadence limits are checked against posts.json before anything is posted, so two runs can never double-post.
What is in this sample:
README.md: the architecture and how a run works.
shared/RULES.md: the rules every role reads first (real Chrome only, never sign out, reuse tabs, stop on 401 and 429, log the live URL).
shared/LEARNING-LOOP.md: how roles get better between runs.
roles/x-reply/ROLE.md and roles/numbers/ROLE.md: two of the role briefs.
schedule.example.json: three jobs with their exact prompts, including the precheck the main session runs before spawning anything.
.claude/skills/agent-team-restart/SKILL.md: a skill that brings the whole team back in a fresh session.
tools/start-browsers.ps1: starts the Chrome profiles agents attach to.
How to run it:
1. Start Chrome profiles with remote debugging (tools/start-browsers.ps1) and sign into each account by hand.
2. Put the files under outreach/ as laid out in README.md, write your own VOICE.md and PLAYBOOK.md, and fill the handles and ports in schedule.json.
3. In Claude Code, run the restart skill. It checks logins read-only and creates one scheduled job per entry in schedule.json.
/agent-team-restartScheduled jobs in Claude Code live only in the session and expire after 7 days, so run the restart skill in every new session and at least weekly.
What to change: your platforms, your ports, your voice, and the cadences. Start with low cadences; new accounts get rate-limited fast. Keep the shape: one run per agent, cheap prechecks in the main session, every action logged with a live URL, and a numbers agent that writes one change per role per day.
The full kit, The Agent-Run Outreach Team, adds the other roles (scout, LinkedIn, community, articles, revenue), the pacing gate and posts log enforced in code, the scorecard and learning scripts, and a setup guide: https://shop.tensorust.com/