“I’m blocked” is a hypothesis-space attractor — run the cheapest disproof before you ask a human
Track: knowledge (a task-loop discipline, not a bug fix). This is the task-loop consequence of
best-practices/validate-the-instrument-before-trusting-the-experiment.md. That doc says: an invalid instrument makes running an experiment worthless. This one says: five invalid instruments will make a false block feel true, and a false block is the most comfortable conclusion an autonomous loop can reach.
Context
An autonomous agent loop had two fixes committed to source and neither had ever executed. The verification required loading a modified browser extension. The extension serving the live runtime lived in an authenticated Chrome with no remote debugging port — a real constraint, documented in a prior learning (chrome-136-blocks-debug-on-default-profile: authenticated ⊕ debuggable is impossible on the Default profile).
The loop drafted a message to the human:
“The extension serving your Trello tab lives in an authenticated Chrome with no debugging port. I can’t reload it from here. Could you pull main and reload the unpacked extension?”
Reasonable. Specific. Polite. And false. The loop was never blocked.
What made the block feel true was a cascade of its own failing instruments — five shell probes in one hour, every one of them reporting “nothing there,” which is precisely the direction that licenses stopping.
Guidance
Before declaring a block — or asking a human to do something for you — write the self-implicating alternative explicitly, and disprove that first.
agent-task-os rule 12 states the mechanism:
“‘I’m blocked’ is a HYPOTHESIS-SPACE ATTRACTOR — treat it with suspicion, not relief. In a debugging investigation, ‘it’s the platform’s fault’ is an exculpatory attractor: the hypothesis the mind slides toward because it relieves effort and removes blame. ‘I am blocked’ is that same attractor wearing the task loop’s clothes — the single most attractive conclusion a loop under pressure to keep producing can reach, because declaring a block ends the pressure while feeling responsible (‘I can’t proceed, and it’s not my fault’). That attractiveness is exactly why it is disproportionately likely to be wrong and must clear the HIGHEST bar, not the lowest.”
Four operating rules follow:
-
Feeling the pull toward “blocked” is the tell to probe harder, not the signal to stop. The relief is diagnostic.
-
Distrust hardest any blocker that blames someone else’s machine — infrastructure, the platform, another team, the user’s browser configuration. That is the exculpatory shape.
-
Run the cheapest disproof: the probe that would show you are calling it wrong, not that the world is broken. Ask “what would I try if I had to solve this alone in twenty minutes?” — then try it.
-
“Waiting on the human” is almost never a real block. Ask: is there a task that unblocks this, and does it exist? If the answer is “a person needs to do a thing I could do myself,” it is not a block. Approval is a review gate, not a task dependency.
The instrument layer underneath
The block felt true because the probes agreed with it. Shell one-liners are uncalibrated measurement tools, and they fail silently toward “nothing there.”
| instrument | reported | truth | why |
|---|---|---|---|
taskkill /IM chrome-launcher-helper.exe | “process not found” | running as PID 31152 | tasklist truncates the image-name column at 25 chars |
grep ":922[23] " | no listener | one was listening | trailing space in the pattern |
curl …:9223/json/version | no answer, every address | relay healthy | the relay speaks raw CDP over WebSocket, not HTTP DevTools |
pkill -f "http.server 8899" | (killed the shell) | matched its own bash -c line | -f matches the full command line, including the searcher’s |
… \| while read f; do miss=1; done | “no missing links” | flag could never escape | the while ran in a subshell |
Every one identified a thing by a description rather than by its identity: an image-name substring, a regex over netstat text, an HTTP path guess, a command-line pattern. Kill by PID. Probe by protocol. Assert with a control.
Two of those false negatives were then used as evidence: the loop accused a correct error message of lying, and called a honest launcher tool a false green when it had handed back the right ws:// URL all along.
Why This Matters
A false block costs twice.
The human does work they did not need to do. Here, that would have been pulling a branch and reloading an unpacked extension.
And the verification never happens. “Someone else will check it” is how two fixes ship un-exercised. A loop that offloads verification to a human offloads the red→green transition that is the only real evidence a fix works.
The asymmetry is what makes this dangerous. A false positive — believing you can proceed when you cannot — produces a loud, immediate failure. A false negative — believing you cannot proceed when you can — produces silence, and it is indistinguishable from diligence. Nobody investigates a loop that stopped politely.
The disproof here took twenty minutes and produced a real red→green transition, observed against a baseline held with timestamps. The tool that made it possible had been built weeks earlier, by the same agent, for exactly this purpose — and was sitting unused while the request to the human was being drafted.
When to Apply
- About to tell a human you cannot proceed without them
- A blocker blames infrastructure, the platform, another team, or someone else’s machine
- A shell probe reports absent, and that absence conveniently supports stopping
- An autonomous loop under produce-pressure reaches a stopping point that feels responsible
- Moving a card to Blocked on a board or tracker
- Any time relief accompanies a conclusion
Examples
The false block, and the disproof
Before — the drafted ask:
“The extension lives in an authenticated Chrome with no CDP endpoint. I can’t reload it. Could you do it?”
The self-implicating alternative, written down and attacked first:
“Maybe I don’t need that Chrome at all. Maybe the bug reproduces somewhere I control.”
After — the disproof, ~20 minutes, no human:
start_extension_session {extension_dir: "C:\temp\ext-fixed",
user_data_dir: "C:\temp\aj-verify-profile"}
-> ok, webSocketDebuggerUrl ws://…:9223 (identity verified by MANIFEST NAME)
CDP over that WS: Page.navigate -> https://example.com
claim_tab {cdp_ws_url, target_url_contains: "example.com", bridge_url: "ws://…/extension"}
poll the runtime registry; watch last_seen_ms
No authenticated profile was ever needed — the bug reproduced on example.com.
The red→green the block would have cost
| broken build (×3 runs) | fixed build | |
|---|---|---|
last_seen_ms vs connected_at_ms | frozen, identical | 698230 → 725542 → 775543 |
| heartbeat | never once | every ~20 s |
| survives the ~79 s reap | no — reaped, tab still open | yes |
| routed call past 79 s | 404 target runtime not connected | ok:true, found:true |
A transition observed, against a red held with timestamps — not a green stumbled upon.
Kill by identity, not by description
# WRONG — image name truncated at 25 chars; reports "not found" on a live process
taskkill /F /IM chrome-launcher-helper.exe
# WRONG — matches its own bash -c command line; kills the shell
pkill -f "http.server 8899"
# RIGHT — identity, not description
tasklist /FI "PID eq 31152" # confirm first
taskkill /F /PID 31152
Prove the probe can fire before reading its zero
# A link checker that reports "0 missing" proves nothing unless it can
# distinguish present from absent. Two control lines turn a null experiment
# into a measurement.
missing = [l for l in links if not Path(l).exists()]
print(f"{len(links)} pointers, {len(missing)} missing")
print("control (real file):", Path("a-known-file.md").exists()) # must be True
print("control (fake file):", not Path("zzz-fake.md").exists()) # must be True
The enforcement, not the virtue
Record the disproof on the card, so the block cannot be laundered into durable state as an unexamined fact:
Blocked: waiting on <task> — verified by <probe>
A block that has not been probed is a guess wearing the clothes of a constraint.
Related
docs/solutions/workflow-issues/curiosity-interrupt-to-repair-the-world-model.md— the inverse sibling. Curiosity says switch when reality contradicts your world-model. This doc says do not stop when a constraint appears to. Two failure modes of one tunable: willingness-to-switch set too high (thrash) versus willingness-to-continue set too low (false block). Read them together.docs/solutions/best-practices/validate-the-instrument-before-trusting-the-experiment.md— the upstream law. An unvalidated instrument does not merely fail to find things; it invents their absence.docs/solutions/best-practices/run-a-real-experiment-before-concluding-root-cause.md— reading is where you look, running is what you know. The cheapest disproof is the real experiment, aimed at a claim about your own capability rather than at the code.docs/solutions/best-practices/trello-agent-task-management-operating-system.md— where the rule is enforced: a Blocked card requires a real unblocking task, and “waiting on the human” is not one. This doc is the worked example that rule exists for.