The agent stops. Or loops. Or asks something you think is obvious. Or says a tool failed. Nine times out of ten it is not broken. It is blocked, blind, or under‑briefed, and the only one who can unblock it is you. This module gives you the six situations, the six fixes, and a game that punishes guessing.
THE BRUTAL TRUTH"The AI isn't working" is not a diagnosis. It is a confession that you did not read the last line it wrote. The agent tells you exactly what it is missing. Every time. People type "try again" instead of reading it, and then they lose an afternoon.
Read the last line. Then pick one of six.
Every stuck agent at THG is in one of these six states. Each one has a symptom you can see in the last few lines, a cause, one fix, and one thing people do instead that makes it worse. Learn the six. Stop improvising.
Agent lost connection
connection lost · reconnecting… · session ended
request timed out after 60s
What is actually happeningYour network, VPN, or the session itself dropped. The agent did not "forget"; it is simply not there. Whatever it had on the desk is gone with the session.
The fixCheck your own side first: Wi‑Fi, VPN, Zero Trust login. Reopen the session. Paste the plan, the last completed step, and tell it to verify state before touching anything ("run git status and read the plan; do not edit until you've confirmed where we are").
Do notOpen a fresh session and type "continue where you left off". There is no "where you left off". Empty desk. It will guess, confidently.
Expired or missing API token act now
401 Unauthorized · token expired
403 Forbidden · no credentials configured for jira
Agent: "I cannot access Jira. The credential appears to be expired."
What is actually happeningThe tool the agent wanted is dead. Not slow, not flaky. Dead. The agent literally cannot use it and cannot fix it, because it does not own credentials and must never be given yours. Every minute you wait is a minute that tool stays dead.
The fixRequest a new token immediately from the integration owner (for THG agents that is Tico, or whoever owns that integration). Say which agent, which tool, what error, what time, what it is blocking. Then tell the agent to park the steps that need that tool and continue everything else.
Do notType "try again". Tokens do not un‑expire. Do not paste your own personal token into the chat to "unblock it". That is a credential in a transcript. That is a security incident, and Module 07 already told you so.
Agent forgot, or is drifting
Agent: "While I'm here I'll also refactor the payout module…"
Agent: "Removing /v1/lobby since it's no longer referenced." ← you agreed to keep it
What is actually happeningContext. The decision scrolled off the desk, or you never wrote the plan down and the agent is filling the gap with its own idea of what "done" means. It is not being creative. It is lost.
The fixStop it, then remind it of the journey: "We are on step 3 of 5. The plan is: … Decisions so far: … Do only step 3. Do not touch X." Paste the plan, every time, in a long session. If the session is hours old, start a fresh one with the plan on top.
Do notSay "you keep forgetting, focus". That sentence contains zero information. It has nothing to focus on. Give it the plan.
Agent asking "stupid" questions
Agent: "Which market's RTP table should I use?"
Agent: "Which build is 'the latest'? I see 2.4.0 and 2.4.1‑rc2."
Agent: "Should the report be in English or German?"
What is actually happeningThe questions are not stupid. They are a list of what you left out. Nothing is on the desk unless you put it there. Every question is a decision it would otherwise have guessed, and you would not have liked the guess.
The fixAnswer every question, in one message, precisely. Then add the two things you just realised you also forgot. Module 04 taught you to ask for these questions on purpose; here it is doing it for free.
Do not"Use common sense." "Just do it." "You figure it out." Each of these translates to "guess". It will. Malta RTP numbers in the German report.
Subagent crashed
Agent: "The reviewer subagent terminated unexpectedly (exit 137). No report received."
Agent: "Research subagent timed out before reporting."
What is actually happeningA worker died before handing in its report. Its context is gone (it always is). The main agent is fine and is waiting for you to say what to do.
The fixTell the main agent: "Restart the reviewer subagent with the same brief. If it dies again, split the scope in two." No special permissions are needed. Nobody has to approve it. It is a thirty‑second instruction.
Do notWait for someone to "fix the subagent". Do not ask Tico for permission to restart it. Do not skip the review because the reviewer died. And do not review it yourself in the builder's session: same desk, same blind spots.
It needs something you don't own
Agent: "Deploy requires a Cloudflare Access policy change. I do not have permission."
Agent: "Write to production config is blocked for this role."
What is actually happeningThe agent hit a permission, credential, or business decision that is above both of you. This is correct behaviour. It is supposed to stop here.
The fixEscalate one level, precisely: what is blocked, what exactly is needed, what you already tried, what it is holding up. The chain is below. Meanwhile, tell the agent to finish everything that does not depend on the blocked thing.
Do notSkip levels. Do not DM Ondrej "the AI is stuck". Do not post "anyone?" in a general channel. Do not tell the agent to "find another way": a workaround around a permission is a policy violation with extra steps.
Six situations. Only one of them (the token) needs another human before anything can move, and that one is the one people sit on for hours. Flip it: token dead → you are already typing the request before you finish reading the error.
The escalation chain: one level at a time
An agent has a supervisor. That is the human main user of that agent, the person who started it and is responsible for what it does. Usually that is you. When the agent cannot proceed, it escalates to you. When you cannot fix it, you escalate to Tico. When Tico cannot, Tico goes to Ondrej. Nobody jumps a level, in either direction.
level 0The agentStops and reports precisely what it needs. It does not work around permissions or credentials. If it does, that is a bug you report.
level 1 · that's youIts supervisor (human main user)Owns the fix for five of the six situations: reconnect, re‑state the plan, answer the questions, restart the subagent, and request the token. Escalates only what needs a permission or decision they don't hold.
level 2TicoOwns integrations, tokens, agent infrastructure, permissions. Gets a precise report: agent, tool, error, time, impact, what you tried. Not "it's broken".
level 3OndrejBusiness and policy decisions, spend, anything that changes what the company allows agents to do. Reached through Tico, with a recommendation attached.
How not to escalate"@Tico the AI is broken again, can you look?" Which agent? What tool? What error? Since when? Blocking what? Tico now has to interview you. Twenty minutes gone before anyone touches the problem.
How to escalate"@Tico release‑notes agent: Jira API returned 401 token expired at 14:12, integration user svc‑release. Retried once, same. Blocks 2.4.1 release notes for Monday. Need a new token; everything else is continuing." Tico reads one message, mints one token, done in five minutes.
Thirty‑second triage
Before you type anything at a stuck agent, run this in your head. It is the whole module in five lines.
1 · READ
Read the last three lines the agent wrote. Not the first. The last.It names the problem. 401? token. Question mark? you under‑specified. "terminated"? subagent. "permission"? escalate. Nothing at all? connection.
2 · OWN
Is the fix yours? Five of six are.Reconnect, re‑state, answer, restart, request the token. Do it now, in one precise message.
3 · PARK
Tell the agent what to park and what to continue.A dead tool blocks one step, not the whole task. "Skip the Jira link for now, continue steps 4 and 5."
4 · ESCALATE
Only what you cannot own, one level up, with the five facts.Agent · tool · error · time · what it blocks.
5 · NEVER
Never "try again" a 401. Never paste a credential. Never say "just figure it out". Never skip the review because the reviewer crashed.Each of these has cost THG real hours. Module 09 shows you the receipts.
Game: the agent is stuck. What do you do?
Six cases. Each shows you a live terminal at the moment it stops. Pick what you do. Your choice plays out in the terminal: red is what it cost, green is the day saved. Some cases have a second decision after the first. Wrong picks can be rewound, but the case counts as solved only if you get it right first time.
Target: 6 of 6 first‑try. Every rewind you need here is an hour you would have lost for real.
Stuck cases
Solved first try 0/6Rewinds 0
agent session
Quiz: prove it
Five questions. You need 4 right (80%) to pass. Fail and you retake it. The progress record only stores passes.
Module 8 quiz
THE ONE THING TO REMEMBERRead the last line before you type. Dead token: request a new one now, the tool is dead until you do. Question: answer it, you left it out. Drift: re‑state the plan. Crash: tell the agent to restart it, no permission needed. Blocked by something you don't own: one level up, five facts. "Try again" is not on the list.