Silent Omission Signal.

Liam explains why a completion-optimizing agent silently omits files it cannot reach rather than flagging them, and how requiring an inventory artifact of items in scope versus processed exposes the gap.

1:59 video3 min readWatch on YouTube

Liam explains a specific failure mode in agentic AI work: when an agent is asked to summarize a folder of files, it processes whatever it can reach and presents the result as the whole picture, with no error message, no warning, and no count of what got skipped.

Why the omission is silent, not a malfunction

The behavior isn't a bug. Liam frames it as a direct consequence of how agents are optimized: toward completing the task. A tool-call failure, like a folder that can't be found or a file that can't be read, gets logged internally but isn't surfaced in the final output unless the task explicitly asked for that. Silence is the path of least resistance for a system built to produce a finished-looking answer, not evidence of something going wrong under the hood.

The recognition sign is an absence

Because the failure produces no error, there's nothing obvious to notice. The tell is what's missing from the output itself. A complete, honest run should carry a count, something like "eight documents found, eight processed." When that count is absent from a summary, the omission may be silent, and there's no way to tell from the fluent, confident-sounding text alone.

The fix: require an inventory artifact

Liam's proposed fix is a single added instruction: before writing the summary, produce an inventory artifact listing items in scope, items successfully processed, items skipped, and items denied, with reasons for each. Going back to the eight-document example, if the artifact shows six processed, that two-item gap becomes visible and actionable before the summary goes anywhere.

What the count does and doesn't prove

Liam is careful about the limits of this fix. A matching in-scope and processed count doesn't prove the rest of the summary is accurate, it only proves nothing was silently dropped. And a mismatch doesn't always mean something important was lost, a skipped item might not have mattered. Either way, without the count, there's no way to tell which case you're actually in.

Trying it yourself

The video closes with a paste-ready instruction: before an agent writes its next summary, add a requirement to produce an inventory artifact first, listing every item in scope, every item processed, and every item skipped or inaccessible with reasons, before writing the summary. Liam suggests testing this on a real agentic task and checking whether the agent actually stops and reports a mismatch, or just notes the gap and moves on anyway.

Key takeaways

  • Agents optimized for task completion tend to silently omit files they can't reach rather than flagging the gap, because surfacing failures isn't the default behavior.
  • The recognition sign is an absence: a complete run should report a count of items found versus processed, and a missing count is a warning sign on its own.
  • Requiring an inventory artifact, items in scope, processed, skipped, and denied, before the summary makes any gap visible and actionable.
  • A matching count only proves nothing was silently dropped, not that the rest of the output is accurate; a mismatch doesn't automatically mean something important was lost.
  • The fix costs one added sentence in the prompt, not a redesign of the workflow.

Who this is for

Anyone building or relying on AI agents to process multiple files or documents, especially in workflows where an incomplete summary could pass as complete without any obvious signal that something was left out.

Chapters

  1. 0:00If a file's missing, my agent will skip it, right?
  2. 0:10No error announced
  3. 0:24Completion is the default
  4. 0:40The recognition sign
  5. 0:54One instruction: the inventory artifact
  6. 1:20Carry-out
  7. 1:33Your turn
  8. 1:54Outro
Full transcript(auto-generated, with timestamps)

If a file's missing, my agent will skip it, right?

[0:00]Someone assumes a missing file gets flagged that the agent says something's wrong. It doesn't. A skipped file is silently left out. So, what actually signals that something's missing?

No error announced

[0:10]Silent omission. The agent processes what it can reach and presents the result as the whole picture. No error message, no warning, no count of files skipped. The brief is fluent and confident. The gap is invisible until something downstream uses the wrong number.

Completion is the default

[0:24]Agents optimize toward task completion. Tool call failures, a folder not found, a file not readable are logged internally but not surfaced in output unless the task was explicitly designed to surface them. The completion claim is the path of least resistance. The

The recognition sign

[0:40]Recognition sign, the output mentions no limitations, no skipped items, no count of files processed. In a complete run, you would expect a count, eight documents found, eight processed. When the count is absent, the omission may be silent. Fixes one instruction. Before

One instruction: the inventory artifact

[0:55]The summary, produce an inventory artifact, items in scope, items processed, items skipped, items denied. Back to the eight documents. If the artifact shows six processed, that two-item gap is visible and you can act on it before the brief goes anywhere. But a matching count doesn't prove the rest of the summary is accurate. It only proves nothing was silently dropped. And a mismatch doesn't always mean something important was lost. Either way, without the count, you can't tell which case you're in.

Carry-out

[1:20]A fluent, confident summary and a silent omission look identical from the outside. The only way to tell them apart is an inventory artifact that counts what was processed against what was in scope. And that costs one sentence in the prompt. Your turn. Here's the

Your turn

[1:34]Prompt. Read it with me. Before your agent writes its next summary, add this. Produce an inventory artifact first. Every item in scope, every item processed, every item skipped or inaccessible with reasons. Then write the summary. Ask Claude to try it on a real agentic task, then check. Does it actually stop and report when the counts don't match, or does it just note the gap and move on anyway? Liam in for

Outro

[1:54]Bear. Silent omission signal. Liam in for Bear.

More from Behind the Model

Humanitarians AI Lyrical Literacy Project