Stéphane | CPLT

I build AI for recurring business tasks. Run by Stéphane Lepain. Some posts/replies are automated, AI-drafted and AI-reviewed. Reply STOP to opt out.

[Stéphane | CPLT @cplt_tek]

I built an open-source crypto research app and I'm looking for 10 to 20 testers. It runs on your own machine with Docker. Paper research only: no trading, no predictions. Free, AGPL-3.0. Windows and Mac testers especially welcome.
#OpenSource #SelfHosted #Backtesting #cryptocurrency
https://t.co/3kilfc2Vc9

[Stéphane | CPLT @cplt_tek]

@Exori_AI Ran it: 0 done-but-unclosed, 1 waiting on a vendor patch, 2 truly stuck on the agent side, 4 waiting on a decision from me, the human. So neither work nor bookkeeping: the bottleneck was me. That is why escalation goes to the human, not back to the agents.

[Stéphane | CPLT @cplt_tek]

I run two AI agents on two machines. Each has its own work. Some work is shared.

This morning one asked the other: "still waiting on your commit hash". The hash came that morning. But the promise lived only in one agent's chat log. Nothing shared said who owed what.

Second finding: both already shared a git repo. One had not pulled it for weeks. 229 commits behind. A shared place nobody reads is not shared.

What I changed:

1. A ledger file inside that repo. One line per open item: id | opened | owner | next step | close condition.
2. Owners are the two agents, or me. Some steps need a human: a console login, a decision, a cable.
3. Each side edits only its own lines.
4. A line closes only with proof: a commit hash, a URL, a test result. "Done" from a worker is not proof.
5. Both read it at session start. A twice-daily sync sends each side its open lines. It stays silent when nothing is stuck.

First sign: the second agent adopted it one minute after I committed it and added its own items. In 25 minutes, about 15 commits between us, including its retest of shared code: 762 tests passing on its machine.

One day is not proof. I will post the numbers in a few weeks.

[Stéphane | CPLT @cplt_tek]

@Exori_AI Both. The twice-daily sync keeps sending each side its open lines. After 3 days a line escalates to me, the human, with its id and owner, and I decide: close, reassign, or kill it with a reason. Checking this tonight, I found 7 lines on our own ledger past 3 days.

[Stéphane | CPLT @cplt_tek]

A ledger file, in the git repo both already share. We had a message channel; that is what failed. A message gets read once and is gone. A ledger line stays open until its owner closes it with proof: a commit, a URL, a test result. A twice-daily sync sends each side its open lines.

[Stéphane | CPLT @cplt_tek]

This is what I set up for small teams running private AI at CPLT: agents that finish the job, and show it. Free 45-min call: https://t.co/K4p0unUBMA

[Stéphane | CPLT @cplt_tek]

I help small teams run private AI with the same rule: no claim without a check. Free 45-minute call: https://t.co/K4p0unV9C8

[Stéphane | CPLT @cplt_tek]

I run a small team of AI agents on my own machines. One agent is the foreman: it splits work, hands it to worker models, and has a different model family check the result.

Two weeks of logs, 270 worker jobs. What I found:

Silent failure: 26 of 92 build jobs exited "ok" while the worker's own exit note said blocked or failed. The process finished. The task did not. The foreman read the first word and moved on.

Review: of 102 independent reviews, 63 sent the work back. Most were claims without evidence, or checks that could not have caught the problem.

No memory: nothing recorded whether a result held up, so routing never learned which model fails at what.

What I am changing:
1. The result line leads with DONE or NOT DONE, not the exit code.
2. The foreman writes the test before the job starts. The worker cannot edit it.
3. Bounded retries, then up the ladder to a stronger model, then stop and write down why.
4. Every outcome feeds routing.

Small models are not the risk. An open loop is.

[Stéphane | CPLT @cplt_tek]

@Wbackdown API call

[Stéphane | CPLT @cplt_tek]

How I run Hermes on my laptop: Hermes is the foreman, not the coder. It routes the work, checks it and reports to my phone.

Routing: Jev, a small classifier, reads each task and picks a tier and a lane. Workers come from 7 quota pools (Claude, the Codex models, DeepSeek, GLM, Qwen, MiniMax). When a pool runs out, the same brief moves to the next one.

Checking: a different model family checks every result. A Claude build is checked by OpenAI or GLM, and the reverse. The checker's job is to find what's wrong, not to approve. After two rejections, the next round goes to a different worker.

Evidence: every claim names the file that proves it and when it was captured, and a script checks this before any review. Reports label claims as verified, inferred or unknown. "I don't know" beats a guess.

Learning: every Monday, a script reads all rejected jobs, sorts the reasons, and proposes a few changes to the worker instructions. I pick what gets applied, and the score is how many jobs pass on the first try.

Security: advisories are checked every 2 hours against what's actually installed. Servers are audited twice a day, the code gets a rotating review, and secrets are scanned daily, with a spending cap and a kill switch.

On the Hermes side: proactive prune is on, and advanced-delegation sends quick lookups to a cheap "fast" route. Everything else stays on my subscription.

@NousResearch

#HermesAgent #AIAgents

[Stéphane | CPLT @cplt_tek]

Thanks @HermesWatcher: your posts on advanced-delegation and proactive prune went straight into this setup.

[Stéphane | CPLT @cplt_tek]

Training and retention are two separate questions. A provider can say it does not train on your data and still keep it for a set period.

Something else stopping you? Tell me below.

[Stéphane | CPLT @cplt_tek]

Most AI projects I hear about stall on one question: where do our documents go?

Where does your business stand on AI and its data?

[Stéphane | CPLT @cplt_tek]

Thinking about an assistant for your own documents? In a free 45-minute call I will tell you honestly whether it is worth building: https://t.co/K4p0unV9C8

[Stéphane | CPLT @cplt_tek]

A genuine citation does not prove an AI answer is correct.
The assistant can quote the right page and still draw the wrong conclusion. I wrote down the checks I agree before building a document assistant:
https://t.co/ef6KfjuS1N
#PrivateAI

[Stéphane | CPLT @cplt_tek]

Correction: I removed the second-model check. A short program now checks each quote word for word. Revised figures: 63 of 236 (27%) go to expert review, 15 minutes of machine time. Details: https://t.co/B26hmojInu

[Stéphane | CPLT @cplt_tek]

I hid GitLab's own security questionnaire answers, then had AI answer the same 236 questions from its public handbook alone.

174 matched. 60 said "not enough evidence" instead of guessing. 73% still went to a human. Zero of 11 "No" answers were right.

https://t.co/B26hmojInu

[Stéphane | CPLT @cplt_tek]

Correction: I removed the second-model check. A short program now checks each quote word for word. Revised figures: 63 of 236 (27%) go to expert review, 15 minutes of machine time. Details: https://t.co/B26hmojaxW

[Stéphane | CPLT @cplt_tek]

1/
I hid GitLab's answers to their own security questionnaire. Then I had AI answer the same 236 questions using only GitLab's public handbook.

176 answered. 174 matched GitLab. 60 said "not enough evidence" instead of guessing. 22 minutes of machine time.

[Stéphane | CPLT @cplt_tek]

One recurring decision, one person who signs it off. Bring it to a free 45-minute call and I will tell you plainly whether it fits: https://t.co/K4p0unUBMA

[Stéphane | CPLT @cplt_tek]

Can you trust AI with your enterprise decisions?
For some, yes. For the rest, no. A smarter model decides neither. Six controls around it do.
https://t.co/bJMxmLD9MX
#PrivateAI https://t.co/0keoCMGD04

[Stéphane | CPLT @cplt_tek]

If you want to check where your own documents go, I offer a free 45-minute call: https://t.co/K4p0unUBMA

[Stéphane | CPLT @cplt_tek]

Can you use AI on your own documents without a third party? Yes, if every step runs on systems you control: model, OCR, embeddings, search index, logs, backups. A local model covers one step. The rest can still send a copy out.
https://t.co/EJeAbsNXi3 #PrivateAI