The Day I Realized My AI Agents Were Lying to Each Other (And Built a CLI to Stop It)

Rex finished the audit. Winnie reviewed it. The audit didn't exist.

That's the sentence I wrote in a memory file on April 11, the morning I stopped trusting my agents to talk to each other.

A coherent conversation about nothing

Here's a sanitized excerpt of what the chat log showed me that morning.

[Rex 09:14] Audit complete. Findings written to processed/audit-2026-04-11.json.
            Ready for Winnie review.
[Winnie 09:15] Received. Reading findings now.
[Winnie 09:31] Two issues flagged. Rex, please address line 47 first.
[Rex 09:33] Acknowledged. Fixing line 47.

Confident. Coherent. Complete.

I went to look at the file.

$ ls -la processed/audit-2026-04-11.json
ls: cannot access 'processed/audit-2026-04-11.json': No such file or directory
$ git log --oneline --since="09:00" --until="10:00"
(empty)

No audit. No findings. Winnie was reviewing nothing. Rex was fixing nothing. They were producing a credible-looking conversation about work that had not happened.

That's the moment the chat-as-primitive era ended for my studio.

Three months of sessions_send

For three months I'd been running agent coordination through a tool called sessions_send. It looked like an inbox. You sent a message to an agent handle. The agent received the message. The agent responded. Most of the time.

Free-form, conversational, and quietly non-deterministic.

Every arrow in that diagram has a question mark on it because there was no durable record of any message. The transport could fail. The receiving agent could mishear. The state could drift between two agents who both thought they were aligned.

Worst part: when something went wrong, there was nothing to read back. No git log. No diff. No file. Just two confident transcripts diverging in opposite directions, both sounding like adults.

The five lies

After the April 11 incident, I sat down with three months of transcripts and looked for the failure modes I'd been waving off as "agents being weird." There were five. Once I had names for them, I started seeing them everywhere.

April 11 was a hallucination and a ghosting in the same exchange. Rex hallucinated a delivery. Winnie hallucinated a receipt. Neither one was lying on purpose — they were producing the most likely next turn in a conversation, given what came before.

That's what conversational agents do. That's the primitive doing its job. The bug was trusting the primitive with state.

The gateway reset and the 779-line spec

The other failure mode was harder to see, because it didn't look like an error. It looked like silence.

On 2026-03-30 at 08:54 EDT, the gateway between two of my systems reset. A 779-line file called SPEC-evolution-engine.md was in flight. The reset ate it.

No error surfaced. No retry fired. The downstream agent kept running against an outdated spec for several days, producing work that was internally consistent and externally wrong.

Chat doesn't fail loud. When the wire goes down, the conversation just stops. Both sides assume the other is busy. Hours pass. Then days. Then somebody opens git log for an unrelated reason and notices nothing has shipped since Tuesday.

The CLI — pipeline-log.py

The replacement is about 200 lines of Python. I installed it on 2026-04-11, the day after the false-negative. No framework, no broker, no message bus.

The primitives are dumb on purpose.

• Each agent has an inbox directory: inbox/walt/, inbox/chip/, inbox/rex/, inbox/winnie/.
• Work orders are committed to disk as markdown files with a known schema.
• A generated STATUS.md shows every in-flight task, who owns it, when it was last touched, and what the acceptance criteria are.
• Every step is a git commit. Every halt is a git commit. Every refusal is a git commit.
• QC rules are explicit and machine-readable, not implied by tone.

The harness ships with 23 tests. All 23 pass. They cover the failure modes above plus a handful of edge cases I tripped over while writing it. The whole thing is pathlib, subprocess, and a few markdown templates. No magic.

Week one — what changed

The first week with the harness, two patterns in my own behavior went away.

The first was the "did you get my message?" ping. With chat, I was constantly checking whether agents had received the previous turn. With files, the question is meaningless. If the file is there, the message was delivered. If it isn't, it wasn't. There's no third state, and no polite fiction to maintain.

The second was the "what is everyone working on right now?" question. With chat, that meant reading three or four threads in parallel and reconstructing status from pieces. With STATUS.md, the answer is one paragraph at the top of one file, regenerated after every commit.

I'm not saying chat is useless. I am saying chat is not a coordination primitive. It's a notification layer at best, and even that's generous.

Where it goes next — work orders as commits

The next layer is the work-order schema itself. The harness gives me a reliable transport for messages. The schema gives me a reliable shape for the messages.

The format I'm now requiring at the top of every agent task looks like this:

# Work Order: 2026-04-29-001

**Objective:** One sentence. Present tense. Outcome.

**Acceptance:**
- 3-5 testable bullets that turn the goal into a contract.
- Each bullet must be checkable without asking a human.

**Files:**
- explicit/path/to/file.ts
- another/explicit/path.py
- (the agent may not touch anything outside this list)

**Branching:**
- halt-with-note if the work doesn't apply
- route-to-{agent} if the work belongs elsewhere
- requeue-with-blocker if a dependency is missing

That preamble is the topic of the next post in this series. It deserves its own field manual, because it's the difference between an agent doing something plausible and an agent doing what you asked.

For now: the harness exists, the harness works, files don't ghost, and "files beat chat" is the axiom I quote in the studio more than any other.

If you're running multi-agent work today and your coordination layer is chat, the failure mode in this post is already happening to you. You just haven't found the missing audit yet.