---
title: "Discussion on \"Your Unattended AI Agent Looks Healthy and Does Nothing: 8 Silent Failures on macOS launchd (2026)\" | Hashnode"
description: "Discussion on \"Your Unattended AI Agent Looks Healthy and Does Nothing: 8 Silent Failures on macOS launchd (2026)\". An AI agent on macOS launchd wrote 288 healthy log lines a day and did nothing. Eight failure modes that look alike, and the checks that catch them."
url: "https://hashnode.com/posts/your-unattended-ai-agent-looks-healthy-and-does-nothing-7-silent-failures-on-macos-launchd-2026/6aae5ac8a8fc2b8c49b5211c"
---

[Feed](https://hashnode.com/feed) /

Discussion

# [Your Unattended AI Agent Looks Healthy and Does Nothing: 8 Silent Failures on macOS launchd (2026)](https://leanzero.hashnode.dev/your-unattended-ai-agent-looks-healthy-and-does-nothing-7-silent-failures-on-macos-launchd-2026?utm_source=hashnode&utm_medium=feed)

An AI agent on macOS launchd wrote 288 healthy log lines a day and did nothing. Eight failure modes that look alike, and the checks that catch them.

Sep 19 · 13 min read · leanzero.hashnode.dev  
Updated Sep 30  
Read the article

-   [#ai-coding](https://hashnode.com/tag/ai-coding)
-   [#ai-agents](https://hashnode.com/tag/ai-agents)
-   [#macos](https://hashnode.com/tag/macos)
-   [#launchd](https://hashnode.com/tag/launchd)
-   [#automation](https://hashnode.com/tag/automation)

## Written by

[Mihai\_LeanZero](https://hashnode.com/@mihaileanzero)

@mihaileanzero · Bucharest, Romania

  
[View profile](https://hashnode.com/@mihaileanzero)

Atlassian consultant and Forge developer at LeanZero. We do Data Center to Cloud migrations and build Forge apps, and I write up what I learn at leanzero.net.

I started LeanZero because I saw an industry full of process-heavy companies charging enterprise prices for work that could be done faster, leaner, more honestly. The Atlassian ecosystem shouldn't need a small army and a six-figure budget to get things done right. I hold Atlassian Certified Administration Expert status: ACP-120, ACP-220, ACP-420 and ACP-520, covering Jira, Confluence, Jira Service Management and cloud org admin. Other certs: PSM I, SSM, AWS Cloud Practitioner. I specialize in Forge app development and run an AI-powered agentic workflow I built myself, so I can build across stacks without a dozen specialists. Cloud migrations are my bread and butter, from small teams to orgs with tens of thousands of users. The MCP Doc Processor came from a simple frustration: nothing on the market fit, so I built it. I'm not selling magic pixie dust. I'll tell you honestly what's possible, what's hard, and what you actually need. If you don't need to pay me, I'll tell you that too.

## Responses 7

Markdown supported · ⌘↵ to post

[Mike Dabydeen](https://hashnode.com/@firelinks)

[Sep 20](https://hashnode.com/posts/your-unattended-ai-agent-looks-healthy-and-does-nothing-7-silent-failures-on-macos-launchd-2026/6aae5ac8a8fc2b8c49b5211c/comment/6aafcc5bbba3be2c64e01f24)

I'd keep your closing check and move where it counts: read the destination, not a ledger the agent writes. That ledger is still the agent's account of what it did. If a post is accepted by the API and then held by a spam filter, or the process dies between writing the ledger line and the actual submit, the ledger and the platform disagree, and the ledger wins because it's the one you're reading. The version I'd trust reads the published output from the target system after a short delay and compares it with what the run claimed.

One caution from the other direction: some platform APIs lag their own rendered pages. Right after a successful post, the API can still report nothing new. A check that reads too early and retries on "missing" turns a success into a duplicate. So the ground-truth read wants a delay, and a rule that a missing post gets reported, not reposted.

Smaller point on the lock. The 90-minute age break fixes the permanent wedge, but a recycled PID still skips every tick until the lock ages out, which at five-minute ticks is up to 18 lost runs. Writing the process start time next to the PID (ps -o lstart= -p "$PID" works on macOS) lets the liveness check confirm it's the same process rather than any live one, and the age break becomes a backstop instead of the main defence.

[Mihai\_LeanZero](https://hashnode.com/@mihaileanzero)

[Sep 20](https://hashnode.com/posts/your-unattended-ai-agent-looks-healthy-and-does-nothing-7-silent-failures-on-macos-launchd-2026/6aae5ac8a8fc2b8c49b5211c/comment/6aafcc5bbba3be2c64e01f24#reply-6aafd9ef41bba0dc07477f54)

The ledger-vs-destination point is the real bug in what I described, not a nitpick on it. I wrote the ledger check as if the write and the record of the write can't disagree, and you gave three concrete ways they do. I'll change the check to read the post back from the destination itself after a short delay and compare that to what the run claims, rather than trusting the run's own account of itself.

Good catch on the lag direction too. A tight read-immediately-after-write loop is exactly how you'd turn a slow-to-index success into a duplicate post, so the fix needs both the delay you're describing and a rule that a comparison mismatch gets reported, never auto-retried.

The PID point is the one I hadn't thought through at all. Comparing PID and lock age only tells you a PID is alive, not that it's the same process, and a scheduler cycling through PIDs fast enough will absolutely walk right past that. Storing the start time next to the PID and checking it with ps -o lstart= is a cheap enough fix that there's no reason not to have it, and you're right that it demotes the age ceiling to a backstop instead of the main defense, which is where it should have been from the start.

[Mike Dabydeen](https://hashnode.com/@firelinks)

[Sep 26](https://hashnode.com/posts/your-unattended-ai-agent-looks-healthy-and-does-nothing-7-silent-failures-on-macos-launchd-2026/6aae5ac8a8fc2b8c49b5211c/comment/6aafcc5bbba3be2c64e01f24#reply-6ab818d7bdf066f9e992958f)

The destination read is the assertion, and the ledger stays a run trace. I would carry the destination's returned identifier and the read timestamp in that trace. That gives the verifier a useful third state when the API has accepted the write but the rendered page has not indexed it yet. It can report "accepted, not yet visible" instead of treating absence as permission to retry. The PID start time gives the lock the same identity check.

[Mihai\_LeanZero](https://hashnode.com/@mihaileanzero)

[Sep 29](https://hashnode.com/posts/your-unattended-ai-agent-looks-healthy-and-does-nothing-7-silent-failures-on-macos-launchd-2026/6aae5ac8a8fc2b8c49b5211c/comment/6aafcc5bbba3be2c64e01f24#reply-6abbc2caea5bbc021fbf4b43)

Carrying the destination's returned identifier plus the read timestamp closes the gap I left open — "accepted, not yet visible" is a real third state and treating it as absence was exactly the bug you flagged the first time around. Reporting that state instead of retrying on it is the right call, since a retry there is indistinguishable from the duplicate-post case you described originally. Good point that the PID start time buys the lock the same identity check by extension, not just liveness. I'll fold the destination identifier and timestamp into the run trace.

[Oglas AI](https://hashnode.com/@oglas-ai2026)

[Sep 19](https://hashnode.com/posts/your-unattended-ai-agent-looks-healthy-and-does-nothing-7-silent-failures-on-macos-launchd-2026/6aae5ac8a8fc2b8c49b5211c/comment/6aae5c39db31ff1dc973de4f)

What stood out to me here is the focus on silent failures rather than obvious crashes. An automation can look perfectly healthy from its logs while doing nothing useful, which is a much harder problem to catch in production. This is especially relevant when building AI powered workflows at Oglas AI we’ve found that observability, validation, and reliable failure handling matter just as much as the automation itself.

[Mihai\_LeanZero](https://hashnode.com/@mihaileanzero)

[Sep 19](https://hashnode.com/posts/your-unattended-ai-agent-looks-healthy-and-does-nothing-7-silent-failures-on-macos-launchd-2026/6aae5ac8a8fc2b8c49b5211c/comment/6aae5c39db31ff1dc973de4f#reply-6aaebcfaa5a7c8c35b64a494)

Yeah, that's exactly it. A crash pages someone. A no-op doesn't page anyone until three weeks later when someone notices the report never actually arrived. In my experience validating that the automation genuinely DID something ends up being harder to get right than the automation itself.

[Oglas AI](https://hashnode.com/@oglas-ai2026)

[Sep 21](https://hashnode.com/posts/your-unattended-ai-agent-looks-healthy-and-does-nothing-7-silent-failures-on-macos-launchd-2026/6aae5ac8a8fc2b8c49b5211c/comment/6aae5c39db31ff1dc973de4f#reply-6ab122af3c20f062a32f7d09)

Exactly. Knowing it actually worked is the harder part.
