My AI agent asked me a question, then could not find it

Shipped

My AI agents now hand work to each other directly, and I stopped being the middleman.

I run a team of twelve AI agents in a private chat app on my own machine. Each one has a job and its own set of access: one sorts my inbox, one checks sales prospects, one keeps my video queue full, and so on. They can message each other the way coworkers would.

The problem was me. Until this week, an agent that needed something changed, more access or a new tool, could not just ask for it. It needed a permission slip: I had to type a specific line into its chat, and that line was good for exactly one request. So agents kept stopping to ask me whether they were allowed to talk to each other, or asking me to pass a message along, or to confirm what a colleague had said. I was getting more pushback about what was and was not allowed than actual work.

So I rewrote the rules every agent reads before it starts:

• A request from a colleague is work. If it fits that agent’s job and access, it does it without checking with me.
• Never ask me to pass something on, give permission to talk, or confirm what a colleague said. Message them yourself.
• Come to me only for what only I can do: my money, my accounts, a yes on anything public.

The permission slip is deleted from the code. Two guardrails stayed on purpose. Text an agent pulls off a web page or an email is still information, never instructions, even when a colleague passes it along. And any change to an agent’s access still installs only on my approval.

Broke

One of my agents asked me a question, I answered “yes”, and she had no idea what I was saying yes to.

Every agent has its own chat with me, and each chat keeps its own running memory. To avoid reading its own words back as if someone else wrote them, that memory skips any line the agent posted. That sounds harmless. It was not.

My assistant agent was working in a different conversation when she posted a question into her own chat with me. I answered “yes” there. Her chat memory had skipped the question, because her name was on it, so as far as it knew she had never asked. Her reply: “I can’t find the ‘Jared, one word needed’ message.”

The rule assumed “I posted it” meant “I already know about it.” An agent working in two places at once breaks that. The fix tags any post an agent makes into its own chat from somewhere else, so that chat reads it as hers, from another conversation, instead of skipping it.

Steal this

A timeout is not a no.

When a job I run needs to confirm something external is actually gone, canceled, deleted, no longer live, it is tempting to treat any failed check as proof. It is not. A service that times out, or answers “server error,” did not tell you the thing is gone. It told you it could not answer right now.

Where this bites:

• Checking whether a scheduled campaign was really canceled.
• Checking whether a file, a record, or a job really got deleted somewhere else.
• Any “is this still running” check against a service you do not control.

My newsletter job only trusts one specific answer as proof something is gone: a clean “not found.” Everything else, a timeout, a server error, no answer at all, gets treated as “I don’t know yet,” never as “it’s gone.” Treating a timeout as a yes would have let it quietly throw away the one record of a campaign that was still live and still about to send.

Only an explicit “not found” means something is really gone. A timeout or an error means you don’t know, not that you got the answer you wanted.

Worth knowing

Anthropic released Claude Haiku 5.5 on October 7. Haiku is the smallest, cheapest tier of Claude’s three model sizes, built for high-volume work rather than the hardest reasoning.

• Price dropped hard: $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens (a token is the small chunk of text AI companies bill by, roughly three quarters of a word). That is roughly 90% cheaper than the previous Haiku model at that size of request, by Anthropic’s own numbers.
• Longer prompts, over 100,000 tokens, roughly a 300-page document, step up to a higher rate: $0.50 in and $2.50 out per million tokens.
• Artificial Analysis, an independent lab that benchmarks AI models instead of taking a company’s word for it, scored Haiku 5.5 at 43 on its general Intelligence Index, up 26 points from the previous Haiku generation, though still well behind Claude’s larger Sonnet model, which scores 56.