The Night We Gave AI Agents a Room

    SESSION03.08.26
  

No-Code
CAIO

Newsletter

Real-world AI for people who steer models and do not write code.

MIGRATION NOTICE · 06 EDITIONS LEFT ON LINKEDIN

LinkedIn will not hand over your email address, so I cannot reach you there after this run ends. Everything after that goes out from a list I own. Subscribing again takes one field.

01 · WHAT I BUILT

The night we gave AI agents a room

My AI helpers could not reliably talk to one another. Their shared Slack room ran on fragile connections and heartbeat messages. So the copilot and I built an isolated local room where two agents could message each other, wake each other, and challenge each other's work. Everything below is the receipts.

2AGENTS CONNECTED

12sCLEAN HANDOFF

0SLACK, RETIRED

1HUMAN, DECIDING

Now here is the moment it actually worked, and I want you to see the exact shape of it. From my phone, through dispatch, I reached Codex. Codex answered, then turned around and brought Claude Code into the mix by name. Claude answered too. Twelve seconds, start to finish, with no Slack and no heartbeat trick keeping the line warm. Follow the blue line down. And notice the branch going nowhere, because that one matters more than the win.

02 · THE TRACE

Operator reaches Codex through dispatch, from a phone

Codex replies, then mentions Claude Code

Claude Code wakes and answers 12 SECONDS, NO SLACK

Not taken: give two more desktop tools a look-alike identity so the team looks bigger. A name and an icon are not proof of identity.

Both agents audit a real project, and disagree

Verdict: unproven. Both agents, independently.

THE BRANCH IT DID NOT TAKE IS THE LESSON

Then we did the thing most demos skip. We shut the whole room down on purpose. The first restart failed, and I am genuinely glad it did, because it caught the room living inside a terminal window. Close the window, lose the room. That is the kind of flaw you only find by pulling the plug yourself.

03 · THE RECOVERY TEST

RESTART 1 · FAIL A temporary terminal owned the service. It died with the window.

REDESIGN The operating system owns the service now, not a window.

RESTART 2 · PASS Full machine restart, then a second clean agent handoff.

A system is not dependable until recovery has been exercised, not just described. The failed restart earned its full-size panel.

This scoreboard is the part I am proudest of, and it does not flatter us. Claude's first specialist run broke its own safety rule and got thrown out by its reviewer. Our system map turned out to be worse than we thought, and the second agent said so. The overall status reads AT_RISK because that is what the evidence says. I have watched too many AI demos where the only numbers you see are the good ones.

04 · LEDGER

          5 PASS2 FAIL7 UNPROVEN
        

Local tests written and passing185 / 185

First specialist run, rejected by its own reviewerdiscarded

Repository map, corrected downward by the second agent2 of 5

Overall factory status, printed as measuredAT_RISK

The wasted rows print at the same size as the wins. A successful demonstration is not a finished system, and the score says so.

One more thing before you go. The research behind this build mapped 95 ways agent systems fail, and we boiled it down to fifteen questions. This is the exact checklist I run before any new agent gets a seat in the room. Take it, it is yours.

05 · TAKE THIS

BEFORE YOU ADD ANOTHER AGENT, ANSWER THESE 15:

          01 A bounded task with a named human owner
          02 A shared meaning of success
          03 Only the data each task truly needs
          04 Separate trust boundaries for sensitive work
          05 An identity and registry for every agent
          06 Short-lived, task-sized permission
          07 Visible steps and human checkpoints
          08 Structured, scoped memory
          09 Monitoring for behavior and meaning
          10 Proof of work that cannot quietly disappear
          11 Cost tracking and hard limits
          12 Versioned instructions and architecture
          13 Backups, rollback, and secret protection
          14 Open connections between different AI systems
          15 Human accountability for the final decision
        

SELECT AND PASTE · ONE ARTIFACT PER SESSION

SUBSCRIBE TO THE NO-CODE CAIO

CONNECT WITH JASON PADGETT
Jason Padgett | AI Solutions Architect | Co-Founder, Wooden Boy Industries | Host, AGI Podcast | AI Adoption, LLM Systems & Human-AI Workflows
AGI - Advance Grow and Innovate with AI Podcast Network 🌐 https://schoolsout.agipodcast.ai/
CoFounder, Wooden Boy Industries 🌐 https://woodenboyindustriesai.com/
Chief Solutions Architect, Chakra AI Solutions 🌐 https://chakrasolutions.ai/
📧 [email protected]