# How I Built My Own AgentOS on Claude's Agent SDK (So You Can Too)

Danny Postma walks through the Agent OS he built over six months on Claude's managed agents / Agent SDK: least-privilege throwaway containers, a spec-to-PR template pipeline, an inbox instead of a terminal, and open-ended goal loops with spend caps.

Tags: youtube-inbox, source, agents, agentos, claude, danny-postma, workforce

## Summary

After six months of leaving Claude Code in a laptop terminal, Postma now claims ~95% of his programming (and most of his businesses) run through a custom UI on Anthropic's managed-agent SDK. Each job boots a unique container, pulls the repo, does one role-walled task, commits, and is thrown away. Humans only show up for approvals and inbox questions; a feature template can run spec → plan → multi-agent plan review → implement → code review → wiki update → PR while he is away.

## Key points

- **Why**: Terminal Claude Code required an open laptop. He wanted overload-and-return, cron, and webhook triggers. Month one was him programming with Claude Code; after that the system mostly builds itself.
- **Least privilege**: Every agent has its own prompt, skills, MCPs, and environment. Plan agent gets plan-mode + Agent OS MCP, not GitHub. Customer-support bot gets Front MCP, never Gmail or repos. Environments can restrict networking (e.g. only api.front.com). Secrets injected per session and stored in Google's token encryption (name forgotten on camera).
- **Filesystem**: Sessions are ephemeral, so persistent files live on Cloudflare R2 behind an MCP. Agents get a folder; writes/deletes go through server-side checks so a model cannot wipe storage.
- **Task pipeline (Kanban)**: todo → doing → review → done. Tasks can run once, on a schedule, or recurring. The "compound engineer" / feature template is a chain of follow-up tasks, some gated on human approval: write spec (must approve) → plan agent → review coordinator (feasibility, scope guardian, coherence; consolidates must-fix / should-fix) → revise plan → implement → code-review coordinator → senior dev applies fixes → librarian updates an internal wiki → human reviews the PR. Example: spawned 3pm, done ~9pm (~5–6h) with E2E tests in the loop; he says 99% of PRs just work.
- **Inbox, not terminal**: Agents contact him only when stuck via an inbox MCP (chat, multiple-choice). Mobile PWA push while at the gym. He can live-watch any session's tool calls.
- **Triggers / automations**: Hedgehog Pro customer-support webhook assigns a support rep or AE (cited ~600 fires). Bug-report trigger: diagnostic agent with repo + chat access writes a cause report; on approve it runs the same plan/fix/E2E pipeline. Cron example: monthly LinkedIn content bot.
- **Goals (gauntlet loop)**: Open-ended work with a definition-of-done checklist. An orchestrator re-spawns after each session, reads progress logs + remaining checkboxes, and hires the next specialist (plan vs senior dev). Caps: spend (one uncapped night hit ~$1,000), max runtime, and a stuck-iteration limit (~19). Claude managed-agent API was ~$500/day; he added a $10 Hetzner VM running Claude `--dangerously-skip-permissions` and Grok "yolo" as local runners. Planners stay on Fable/Claude; workers on Grok 4.6 when the VM is free.
- **CLI + YAML**: Each project has an AgentOS tree (agents, skills, templates as YAML) with push/pull/sync. Local Claude brainstorms, then CLI creates the remote goal/task. Offered to open-source prompts/skills if people ask; no repo linked yet.

## Quotes / memorable lines

- "I basically automated 95% of all my tasks to this system."
- "You don't want to give any access to your agent that they don't need."
- "After every session, this container is thrown away."
- "I don't have to manage them, I don't have to wait for them."
- "The only thing I have to do from that PR… is check it out."
- "I run one goal without spend cap one night and it [ran] up to $1,000."
- "The point isn't to copy my tool. It's to build your own version based on these patterns." (video description)

## Sources

- YouTube: https://www.youtube.com/watch?v=Tos-zPxYPuc
- Raw ingest: `raw/youtube-inbox/Tos-zPxYPuc/` (metadata.json, transcript.txt, notes.md, source.url)
- Channel: Danny Postma https://www.youtube.com/@dannypostmaa
- Transcript: Supadata v1 mode=native text=true, lang=en, 22845 chars. ASR spellings are not authoritative.

## Related pages

- [[youtube-ai-agent-workforce]] (Allie K. Miller's 34-agent org vs Postma's role-walled pipeline + inbox)
- [[youtube-hermes-kanban-multi-agent]] (Kanban/SQLite bus + human gate vs his todo/doing/review/done templates)
- [[hermes-agent]] (self-hosted agent OS vs Claude managed-agent SDK + custom UI)
- [[grok-bot]] (managed shared computer vs throwaway per-session containers)
- [[grok-bot-privacy-and-security]] (approvals / least-privilege analogue)