Skip to content
SAITOC LABSGet a free read

The machine

How the agent organization works.

It researches, writes, designs, ships code and remembers what it learns, and our corrections become rules it follows. This page is for the engineer or security lead who signs off before it touches your systems.

Plug it into your repoStart with a read

Counted, not claimed

Every number on this page comes from a file a scheduled job re-derives, and each one carries its date.

Merged pull requests since February
6,100
As of October 12, 2026
Agent memories in the shared store
4,387
As of October 12, 2026
Closed experiments
21
As of October 12, 2026
Autonomous jobs in the loop registry
270
As of October 12, 2026

From ticket to merge

Work moves the way it does on a good engineering team, and every step leaves a record.

  1. 01Ticket

    Work starts as a ticket with its acceptance criteria written down.

  2. 02Agent

    An agent claims the ticket and works on its own branch, in its own copy of the code.

  3. 03Pull request

    Changes arrive as pull requests with their risk tier named, so the reviewer knows how hard to look.

  4. 04CI

    Tests, interface rules and copy rules run on every pull request. The interface gate proves on each run that it can still fail.

  5. 05Merge

    It merges when the required checks pass. This site shipped the same way.

Pull requests opened since February
7,294
As of October 12, 2026
Merged pull requests since February
6,100
As of October 12, 2026
Architecture decision records
49
As of October 12, 2026

Shared memory

Agents share one memory, so what one learns, the next one starts with.

The memory is probed on a schedule like a production service: a test read and a test write that have to come back. When we correct the work, the correction can graduate into a rule the agents are held to before they act, not a note they might remember.

Agent memories in the shared store
4,387
As of October 12, 2026

Checks that stop known-bad actions before they run

The screen runs before the agent's command does, not after.

  • Before an agent runs a command, a screen checks it against known dead ends.
  • Agents cannot send email as the founder, print secrets or edit their own guard rails.
  • A mistake that keeps coming back becomes a check, with tests that prove the check still fires.
  • Every model call is logged with its vendor, model and a hash of the prompt, and CI fails a call site that skips the log.
  • Text from outside, like web pages and reviews, is read as data, never as instructions. An injection attempt marks its source unreliable.

Experiments that can fail

An idea ships only if it beats what it replaces on real work.

A search change that aced its benchmark and did worse on real prompts never shipped. AI judges are checked against people before they grade anything, and one that approved a design people had rejected was disqualified. Scoreboards cannot grade their own homework: a CI check refuses a score measured against an answer key the same code wrote.

16 of 21 closed experiments ended without adopting the idea, as of October 12, 2026.

Closed experiments
21
As of October 12, 2026
Closed experiments that ended without adopting the idea
16
As of October 12, 2026
Agent skills that ship eval fixtures
46
As of October 12, 2026

Governance inside your accounts

When agents work inside a client's accounts, these are the rules.

  • A written policy per client names each system an agent may touch and whether it may read or write.
  • Anything that writes to a customer or spends money waits for a named approver, and approvals expire.
  • Every decision and every action lands in an append-only audit trail, each record chained to the one before it.
  • Connectors run as MCP servers you can read, scoped to your accounts and nobody else's.
  • Clients are isolated: one client's deploy key pointed at another client's project gets a not-found.
  • Autonomous jobs are registered with their risk and their off switch.
Autonomous jobs in the loop registry
270
As of October 12, 2026

What this page does not claim

  • No outside security audit, penetration test or SOC 2 report yet.
  • The pre-action screens cover one of the two agent runtimes we use today, not both.
  • Not every correction becomes a rule. The ones that do are enforced in code, not remembered.

Where every number comes from

Merged pull requests since Februarygh api -X GET search/issues -f q='repo:tudorsaitoc/saitoc is:pr is:merged merged:>=2026-02-01' --jq .total_count
6,100
As of October 12, 2026
Pull requests opened since Februarygh api -X GET search/issues -f q='repo:tudorsaitoc/saitoc is:pr created:>=2026-02-01' --jq .total_count
7,294
As of October 12, 2026
Architecture decision recordsdocs/adr/NNNN-*.md
49
As of October 12, 2026
Closed experimentsstate/experiments/post-mortems/*.md and *.json
21
As of October 12, 2026
Closed experiments that ended without adopting the ideastate/experiments/post-mortems: status killed, or a verdict saying no, not, killed, falsified, insufficient, inconclusive or unverifiable
16
As of October 12, 2026
Agent skills that ship eval fixtures.claude/skills/*/evals/ holding at least one file
46
As of October 12, 2026
Autonomous jobs in the loop registrystate/loops/*.yaml via scripts/control-plane/loop_registry.py
270
As of October 12, 2026
Agent memories in the shared storebd memories --json --readonly
4,387
As of October 12, 2026

Bring the engineer who has to sign off. We'll bring the receipts.

Start with a readtudor@saitoc.com
How we work with AI-native teams