SerSanMulti-agent systemOct – Nov 2026
SerSan Agent OS
The multi-agent system I built for SerSan: agents with roles, a router, hand-offs with a brake that calls in a person, and a Gate that lets an action out only at the level the agent has earned.
- Access
- private repository
Earned autonomy
each agent acts on its own only at the level it has earned
01 What it is
Agents with roles,a Gate on the way out
SerSan's own multi-agent system, which I built for the studio. After a client call, or each morning, its agents prepare call summaries, CRM notes, to-dos, follow-ups, meeting briefs and their own stand-up. A router sends each job where it belongs, the agents hand work to each other on a shared board, and nothing reaches the outside except through the Gate.
Stack
- TypeScript
- Node.js 22
- Claude Agent SDK
- Model Context Protocol
- Anthropic API
- Open-weight models hosted in house
- Zod
- Fastify
- PostgreSQL 16
- Vitest
- GitHub Actions
02 How a job travels
From an eventto the Gate
A job starts from one of 12 kinds of event and ends at the Gate.
- 01An eventafter a client call, or each morning
- 02The routerto a code rule, a model hosted in house, the Claude API, an Agent SDK session, Claude Code or a person
- 03The agentseach with its role, on three lanes
- 04The boardthe hand-offs, and a brake that calls in a person
- 05The Gatethe only part that acts outside
Then the action
- Is done by a person
- Becomes a card a person approves
- Waits in a timed window, with a Stop button
- Goes out on its own, within daily caps
Four levels of autonomy Set for each agent, action type and kind of recipient.
03 Tested
Broken on purpose,piece by piece
Tests that break each piece on purpose, to show what happens when a part fails. It runs on development machines, with test data and a scripted model in place of the real ones.
2,513
Automated tests passed
And 0 failed, on 20 November 2026; there were 777 when the first milestone closed. The suite must give the same names, outcomes and counts in three runs in a row.
28
Failure scenarios, end to end
Plus six continuity tests. The Gate down, two approvals at the same instant, instructions hidden in a third party's text, the emergency stop, the engine dying mid-chain, three clients at once: each ends by checking that no output went above its level.
12
Defects confirmed by adversarial review
In the first milestone: 1 critical, 4 high, 7 medium, each fixed with a test seen failing first. In the second, 17 and 13, none critical or high.
11
Days, first commit to last
304 commits, 82 pull requests, about 96,000 lines of TypeScript (37,600 of source, 58,600 of tests), written by directing one coordinating and two executing Claude Code sessions under my written directives.
04 Guardrails
Autonomyis earned
Written in the code
- 01Each agent's level holds for one action type and one kind of recipient: 9 action types, 10 classes of recipient.
- 02From a card to a timed window: 5 approvals in a row and 30 days without a negative signal. To going out alone: 15 clean windows and 60 days. Always with the owner's yes; a negative signal lowers the level by itself.
- 03Payments are always manual; deletions, credentials, secrets and permissions are off limits.
- 04A content check before anything leaves without a person's approval, and a line in every message sent without a person saying that an AI prepared it.
- 05A register that can only grow, enforced by triggers in the database, and an emergency stop.
05 The record
- Studio
- SerSanSersan LimitedLondonRemote
- Role
- Idea, proposal and direction; built by directing Claude Code sessions
- Period
- Oct – Nov 2026
- Access
- Private, the studio's internal system
- Status
- Built and tested, not in production
- Photograph
- none
More work
- 01SerSan Agent OSSerSan · multi-agent systemThis case
- 02SphereNodeSerSan · live at spherenode.comNext
- 03Sphere AISerSan · RAG mentorOpen
- 04Maintech OMASerSan · workshop ERPOpen
- 05sersan.ioSerSan · Jun – Sep 2026Open
- 06Domus TuaSerSan · real-estate brand platformOpen
- 07Brand filmSerSan · identity and filmOpen
- 08OblaPersonal project · multi-agent systemOpen
- 09Badante24hALS MCL Civitanova · freelanceOpen
- 10Voice AI AgentDOIT S.r.l · internshipOpen
- 11Supply ChainUniversity of CamerinoOpen