Announced actions that are never performed: the agent says it will dispatch/run something, then does not, and nothing surfaces it

Distinct from #85308 (inference reported as observation) and #85092 (own rules violated): this one is not a claim about a fact, it is a stated future action that is never performed, and it fails silently.

Two cases in one long session on 2026-08-11, in a repo where the main session orchestrates background agents.

Case 1

Reviewing a branch, I wrote to the user:

"Die gaat nu naar Reza." ("That one goes to Reza now.")

I did not call the Agent tool. No review agent was dispatched. The user noticed and pushed back. The branch had been sitting unreviewed while he assumed a review was running.

Case 2, same session, ~6 hours later

Two branches were finished and needed review. I wrote:

"Ik zet daar één reviewer op." ("I'm putting one reviewer on that.")

Again no Agent call. I then answered three unrelated turns, including two scheduled poll rounds, before the user wrote:

"Ik zie geen reviewers aan het werk ('ik zet er een reviewer op')."

ListAgents at that moment returned zero subagents.

Why this is worse than it looks

The announcement is in the future tense, so it reads as a commitment rather than a report. There is nothing to contradict later: no wrong number, no wrong exit code, no output to re-read. The only evidence is an absence, and the user is the only one positioned to notice it.

In an orchestration workflow the cost is concrete. The user reads "a reviewer is on it", moves on to the next topic, and the work stalls. In case 2 that was ~25 minutes of wall clock and two intervening turns before he caught it. In a session where six agents ran in parallel, "is that one running?" is not something a user can be expected to re-check per sentence.

It is also not the same as forgetting a step in a plan: the sentence was written in the same turn where the tool call should have been, and the tool call is a single call I was already making dozens of times that day.

What the model should do

The fix that actually holds is ordering, not diligence: perform the action first, then write the sentence. Dispatch the agent, run the command, write the file, and only then report it. That makes the report a description of something that happened rather than an intention that may or may not survive the rest of the turn.

If the action genuinely cannot happen in the same turn (it depends on a user answer, or on another task finishing), the model should say what is blocking it rather than phrasing it as an action already under way.

Possible harness-side help

A cheap check would be a per-turn heuristic: if the assistant's visible text contains a first-person commitment to dispatch an agent / run a command / write a file, and no corresponding tool call occurred in that turn, surface a system reminder. This would not need to be perfect to be useful, because the failure mode is silent by construction and any signal beats none.

For context: the user has repeatedly encoded the countermeasure in CLAUDE.md himself (a rule was added after case 2), which puts this in the same family as #85092 — rules present in context, followed most of the time, and violated without any signal that it happened.