01-projects/personal-brand/drafts

My AI agent didn't crash. It waited.

2026-10-02·post-draft·status: draft-v1

By Ben Wilson, [TITLE: Principal Consultant at phData?]

My AI agent's worst failures never threw an error. It waited for a password prompt nobody saw. It ran for a day and a half with most of its tools quietly removed. It told me four times that a task was blocked on me, when I had finished my part days earlier.

None of these were model failures. They were operations failures, the kind your data platform team already knows how to catch.


The setup

I run an always-on agent on a Mac mini at home: Claude Code, Anthropic's command-line coding agent, running scheduled jobs such as a morning brief and a daily research scan. I control it from my phone. Sub-agents do the heavy work, and a separate critic sub-agent checks the result before I see it.

[BEN: one or two sentences on why you built it and what you use it for most. Keep it personal; no client or employer work.]

1. A credential that asks a human is a stall

On October 2, the agent restarted its connections to outside tools. Some tool connections lacked credentials and hit an approval prompt meant for a human. Nobody was watching, so those tools never came up. Now they fail with an error instead of waiting.

For your agents: an agent's identity must be non-interactive and scoped. On Snowflake, that is now a first-class object: the SERVICE_AGENT user type (generally available since July 2026) gives an autonomous agent its own identity, authenticated by workload identity federation or another non-interactive method. If your agent borrows a person's login, the first multi-factor prompt stops it.

2. Silent stalls look like success

My daily research scan timed out on October 1 and again on October 2. On the first day, it waited on background sub-agents and a wake-up timer that don't behave the same way when the agent runs unattended (headless mode). On the second, the model simply stopped responding after a tool call. Neither run crashed. Each one sat there until the scheduler's time limit killed it.

The job now does its reading inline, with no background agents. Not built yet: a watchdog that kills a job after about ten minutes with no progress in its log.

For your agents: "the process is alive" is not a health check. Measure progress: rows written, steps logged, the age of the last output. You already do this for pipelines with freshness checks. Do it for agents.

3. Your configuration has a supply chain

From September 17 to 18, every new session came up with most of its tools missing. I restarted and reloaded the agent about a dozen times over roughly 35 hours. The cause was a third-party plugin that synced into my account. Its settings file named a default agent, and that agent's tool list was short. Claude Code's plugin system allows exactly this: a plugin's settings.json can set the agent that applies while the plugin is enabled, and an agent's tools field is an allowlist. So every session quietly became that narrow agent.

Another restart didn't find it. A probe with a clean configuration, compared against my normal settings, isolated it. The fix was one command to disable the plugin.

For your agents: anything that syncs into an agent's environment (plugins, prompts, tool configs, shared skills) is a dependency. Inventory it, review it before install, and know who can push to it.

4. A missing permission can take down the control channel

On September 20, the channel I use to reach the agent from my phone went dark. An update changed the path of the agent's program, and a macOS privacy permission no longer recognized it. Only a person can grant that setting. [BEN: how long the channel was down, in your words.]

For your agents: decide in advance what the agent does when its control channel fails, and test that path. Mine had one fallback for sending a message, reserved for real blockers.

5. Verify the blocker at its source

In late July, my agent reported a task as "blocked on Ben's login" four times across two days, including in a morning brief. I had logged in days earlier. A note written about 14 hours after my login said the login was still needed, and every later check quoted the note instead of checking the login.

The rule now: if a blocker is something you can check (an auth file, a deploy, a ticket), check it. Don't quote a note about it. A note is a claim with a date. I wrote it down as a standing instruction.

For your agents: status reports from an agent need the same lineage you'd demand from a dashboard. Where did this "blocked" come from, and when was it last true?

What held up: independent critics

The pattern that caught the most was a critic sub-agent that gets only the finished work and the rules, with none of the producer's reasoning. Today, for example, a critic returned "iterate" on a research note with ten fixes before I ever read it.

I also tested whether a critic from a different model vendor adds anything. On a small benchmark (ten artifacts with planted defects), three models from three vendors each caught all ten. Switching vendors added nothing measurable at that sample size, and with every model at 10/10, the test was too easy to say more.

For your agents: the agent that produced the work should not be the one that approves it. Build the check in as a separate step with its own inputs.

The pattern

Every failure here has a data-platform twin: service accounts, freshness checks, dependency review, runbooks, lineage, and separation of duties. If you run a governed data platform, you already have most of the operating model for agents.

[BEN: a Healthcare / Life Sciences line, if you want one. For example (hypothetical): an agent that drafts prior-authorization summaries overnight fails exactly like #2. It doesn't crash; it just isn't done at 7 a.m.]

Your turn

Which of these five would your current agent setup catch today, and which would you find out about from a user?


LinkedIn cut (~150 words, no link in body)

My AI agent's worst failures never threw an error.

It waited on a password prompt nobody saw. It ran 35 hours with most of its tools quietly removed. It told me four times a task was blocked on me, days after I'd done my part.

None were model failures. All were operations failures:

  1. A credential that asks a human is a stall. Give agents their own non-interactive identity.
  2. "The process is alive" is not a health check. Measure progress.
  3. Your agent's configuration has a supply chain. Inventory what syncs into it.
  4. Plan for the control channel going dark.
  5. Verify a blocker at its source. A note is a claim with a date.

What held up: a critic sub-agent that sees only the output and the rules.

If you run a governed data platform, you already own most of this playbook.

Which of these would your agents catch today?


Ben to decide

  1. Title. Retitled from "What broke when I ran AI agents for 30 days" (incidents span Sep 17-Oct 2, plus one from late July). Alternative in frontmatter.
  2. Naming tools. Draft names Claude Code and macOS only; credential tooling and the control-channel app stay generic.
  3. Home setup detail. Is "Mac mini at home, controlled from my phone" more personal detail than you want public?
  4. Length of #4. The macOS permission outage is the most Mac-specific item. Keep, shorten to two lines, or cut to four lessons?
  5. Benchmark claim. Now limited to "switching vendors added nothing measurable at n=10" (ceiling: all 10/10; no arm tested a critic that saw producer context). Keep or cut.
  6. HCLS line. Add the labeled hypothetical, or leave this post as pure build log.
  7. Snowflake tie-in in #1. SERVICE_AGENT is the one Snowflake mention. Keep it, or keep this post vendor-light.
  8. Byline title confirmation.

Verified product claims (checked 2026-10-02)

Claim Source
Claude Code plugin settings.json can set agent (and subagentStatusLine), applied while the plugin is enabled https://code.claude.com/docs/en/plugins-reference
A sub-agent "runs in its own context window"; the tools field is an allowlist https://code.claude.com/docs/en/sub-agents
SERVICE_AGENT user type is generally available (Jul 23, 2026); authenticates via workload identity federation or another non-interactive method https://docs.snowflake.com/en/release-notes/2026/other/2026-07-23-service-agent-user-type ; https://docs.snowflake.com/en/user-guide/agent-identity

Unverified, cut or check

Deliberately left out

Phone numbers, chat IDs, file paths, credential storage and loading details, the plugin's name, author and origin, internal URLs, cost and spend-limit details, the August repeat of the stale-blocker incident (client work), and the Discord retirement.