The lab
A harness that runs my own agents on my own notes. Separate from Atlas, on my own time, with my own stack.
Anthony Carl · separate from Atlas · 2026 → present
One person, many agents, and the memory they share. The lab exists so that nothing I learn leaves when a project ends, and so that the agents who help me build the next thing read what the last one wrote.
The lab, shown · 1 of 4simulation · invented requests · counts from 2026-09-30
Several agents, and each one started from nothing.
agent one
- ·what the project is
- ·what has been tried
- ·what failed last time
agent two
- ·what the project is
- ·what has been tried
- ·what failed last time
agent three
- ·what the project is
- ·what has been tried
- ·what failed last time
the same context, explained by hand · 0 of 3 times so far
Hire me and the memory comes too.How it is built ↓The desk →
Opened · One person, many agents
I was running several agents by hand, and each one started from nothing.
Without a harness, every session starts with the same explaining: what the project is, what has been tried, what failed. The notes exist, but no agent reads them; the decisions exist, but nobody checks the work before the person does. The cost is not the agents' mistakes. It is being the only memory and the only gate.
The missing piece was not a better model.
Blocked · What breaks without a harness
Without memory the agents repeat; without a gate they ship what reads well.
An agent that cannot read the last post-mortem makes the last mistake again. An agent whose work is judged by another model passes anything plausible. A person who approves everything by hand is the bottleneck the agents were meant to remove. The harness exists to close all three.
So the harness had to be three things at once: a memory, a gate, and a queue.
Reassigned · The decisions
The notes are the database, the checks are the judge, and a person is the last word.
The second brain is the system's database: every request, every verdict and every outcome is written back as a note, so the next request starts from it. Work is judged by real checks, the type-checker, the build, the test run, not by another model reading the output, because a verifiable check is cheaper and cannot be charmed. And nothing ships without a person at the approval queue, because the point was never autonomy; it was leverage with judgment kept.
One request can go to several models, routed by the kind of work, so that cost and quality are a choice per task rather than a global setting.
Then I drew the loop.
In progress · The loop
LAB-0311, code, routed to model A; at the gate.
verdict gatewaiting
- types·
- build·
- tests·
- ·
approval queue
waiting for the gate
the notes
- ✓0309 · playbook · 3 links
- ✓0310 · sent back · written
Scripted simulation of the loop. Invented requests and model names; no code, no project names, no note contents.
A request enters, a model does the work, the gate checks it, a person decides, and the notes remember.
The run beside this is the harness at work, with its internals left out: a request is typed once; the router picks a model by the kind of work; the work comes back through the verdict gate, which runs real checks and sends a failure back once; the approval queue puts a person in front of what passed; and the result, approved or rejected, is written to the notes, which the next request reads first.
One decisionA verdict gate with real checks over an LLM judge. The checks are slower to write and impossible to flatter.
The notes are what compounds.
Closed · What runs today
8,524 notes and 35,748 links, most of them written by the agents.
Counted on 2026-09-30: 8,524 notes, 35,748 links between them, 32 skills and 63 commands the agents can run, and 6,088 notes in the system's own workspace. The counts are the shape of the memory, not its contents, and they are the only numbers this page claims.
No traffic, revenue or user numbers are claimed for anything the lab has made.
A week nowA request typed in the morning is checked, approved and written back before the next one; the second week of a project reads like the second year.
Hire me and the memory comes too.
Next
Do again
Make the memory the database first, and put real checks in front of a person before adding a single agent.
Not again
Trust a model to judge a model. It passes what reads well.
The lab is where the patterns are tried on my own time and my own stack; Atlas is the agency's platform and the two are separate. What they share is the rule: encode the check a person was doing by hand, and keep a person where judgment is needed.
Built with
Hire me
I'd build this for your team.
I am looking for an applied-AI engineering role at a company with real operations to automate. The résumé is one page.