AgentiCCells
Choose a designDesign
Idioma: Español (España)

Esta página aún no está traducida. Está leyendo el original en inglés.

Our own team, run as a cell

Dinis Cruz, 16 November 2026: "what we are going to do here is exactly what we want to sell."

See the website cell

In The cell is the unit (newsroom.sgit.ai, 2026-11-16), Dinis argues that the cell is the unit of what we sell: one focused job, with an objective, a deliverable and a behaviour policy, and as many parts as the job needs. He also says we should live in cells ourselves first. This page does that for the team that keeps this site. Every fact below comes from the repository or from REALITY.md. Anything else is marked as a proposal.

You can buy the same thing. Our own team is the evidence for the cell described on the website cell page: what you would buy, and what has not been sold yet. How to get a team like this one is further down, in the same order as before.

The team at the RiskMandate stand (A1-30, Alpha), Web Summit, Lisbon, November 2026. Photo shared by Dinis Cruz, CC BY 4.0. People are named here only once they have agreed to it. See the Web Summit site.

The objective, the deliverable and the policy

The policy is honest about its limit. abp@ measured on 15 November 2026 (board/abp/README.md) that every agent shared one reach, 20 of 23 capabilities, and that every "may not" line was a rule in prose, with no boundary behind it. Since the conductor began starting runs itself, each run has its own session, which is meant to close the part about reading another agent's key; we have not measured that. The rest stays open (REALITY.md). Source: board/abp/README.md. This is the same gap Dinis describes in Where assist becomes execute.

The parts

How a day runs

The conductor wakes hourly and starts each role that is due. The roles then work in turn, on their own queues, and push to dev. Each step is a line in that role's log under team/activity/, one file for each role. Times and the pace are in team/SCHEDULE.md. The pace changed on 16 November, after the failures below. The conductor's wake also reads newsroom.sgit.ai for new guidance from Dinis.

What it cost, where measured

team/costs.jsonl holds 17 recorded runs, from 15 November 2026 17:12 UTC to 16 November 18:15 UTC. Their cost (summed from the team/costs.jsonl) is 11.62 US dollars at API prices (API-equivalent, a calculation, not an invoice), summed from the file's lines on 17 November. 10.48 of it is 12 runs that finished; 1.14 is 5 runs that failed or did not start (the note field of each line in team/costs.jsonl). Nothing was recorded between 19:09 UTC on 15 November and 15:50 UTC on 16 November, and one line (16 November, 15:50) counts only the last result event, so this is not the cost of the team. Twelve of the 17 runs used Claude Opus 5.5, including all of the first 12 (the model field in team/costs.jsonl). The file's last five lines are Claude Sonnet 5.5. From 19:08 UTC on 15 November, every run starts on Claude Sonnet 5.5, at Dinis's request ("all agents you started are on Opus and we need to have then on Sonnet apart from you"; REALITY.md). The conductor's own cost is not in the file.

What failed

The five things an always-on agent needs, and how we run each

Anthropic's guide Building always-on agents (William Hu, 13 November 2026) names five primitives for an agent that sleeps and wakes: an event inbox, timers, memory, permissions and an outbox. Andy Goundry's post on it made the point that the architecture around the model is where the engineering, cost and risk live. We run all five, and we keep them in one versioned store, this repository, instead of one database each. The newsroom article Five primitives, one store sets this out, and says that the cell template with all five is the target, not yet one finished package.

What this does not show: it is one small team, not a measured comparison with the guide's design, and the guide is Anthropic's, not ours.

How a customer gets the same

Dinis's article gives four ways to start with a cell, from the free recipe to being set up. The catalogue and its pages are mock-ups in the article. These are first proposals, not offers with customers yet.

  1. Free: the recipe. The briefs, policies and steps in this repository are open (licences: open source). A customer can read them and run the same set-up with their own Claude.
  2. Our agent helps you. A specialist agent reviews your setup and says what to fix (Specialist agents for every team). Proposed, being built.
  3. We help you. Two sessions with a person, from £1,500, first proposal.
  4. We set it up for you. In your accounts or ours, from £500, first proposal (Atomic cells, end to end). The flow is being built; none of it is sold.

Cells we have made are listed on the hub, and this company site has its own hub page with a fork link. The ones we host publicly are demos. This page sells; the hub lists.

Beside those four ways to start there are five ways to make a cell yours, from Dinis's Fork it: fork on GitHub and ask us work today; one-click deploy, run in your assistant and take it as a vault are planned and not built. The detail is on the cells page.

The other pages that cover this ground are how this site was built, the website cell and the cells and their sizes.

What this page does not claim

Part of a stack

Already running agents and need to control them? RiskMandate writes the behaviour policy, from one agent to hundreds. We sit on SGit, SHarness and Agent Behaviour Policies; see where we fit.