The signal that holds up: the agent proposes a fix it already reverted. Once the same wrong assumption comes back twice, more steering mostly adds text it keeps rereading. What a fresh session lacks is the list of what was ruled out and why, so that short list beats a full summary. Does /compact count as a reset for you, or as a third option?
Practical rule packs & skills for AI coding tools — Cursor, Claude Code, OpenAI Codex. Writing about agent workflows, project rules, and what actually ships.
For me /compact is a third option. It keeps the session going on a summary of the history, so a wrong assumption that made it into the summary stays. So when it loops, I'd start fresh with your short ruled-out list instead.
Fair point: the summary is where a wrong guess quietly turns into a fact. I'd keep the ruled-out list in a file the fresh session reads first, so it survives the next reset too. Do you keep one list per task, or one running file?
Founder of Aladdin AI, a tool that finds jobs that fit you, writes tailored resumes & cover letters and applies for you. 100 free applications a month, no paywall, no credit card. Building openly.
I build Aladdin, a job agent. My signal is the second repeat of the same fix. One retry can be a bad read. Two means the session is holding a wrong assumption, and steering just adds more words around it. When I restart, I carry over a short summary of what was learned and the verified facts, not the full transcript. The fresh worker then reads the actual page again instead of trusting the old session's picture of it. That fresh read is what breaks the loop.
Practical rule packs & skills for AI coding tools — Cursor, Claude Code, OpenAI Codex. Writing about agent workflows, project rules, and what actually ships.
One catch with carrying over the verified facts: the stuck session is usually the one that verified them, so the wrong assumption can ride along inside the summary. My guess is that's why the fresh read of the page ends up doing most of the work.
Founder of Aladdin AI, a tool that finds jobs that fit you, writes tailored resumes & cover letters and applies for you. 100 free applications a month, no paywall, no credit card. Building openly.
That catch is real. I keep what carries over narrow: only the facts the user gave and what was actually filled in, not the old worker's conclusions. The fresh worker still reads the form itself, and it asks the user when something is unclear instead of trusting the summary. So the summary helps with the starting point, but the fresh read does the checking.
Practical rule packs & skills for AI coding tools — Cursor, Claude Code, OpenAI Codex. Writing about agent workflows, project rules, and what actually ships.
Yonto Player is a media player for your own Jellyfin, Plex and Emby library, on Android TV, phones and tablets. Built with AI coding agents; this is how.
My signal is when it starts undoing its own change from two steps ago. At that point I stop, have it write what it learned into a short notes file (what was tried, what was ruled out and why), and start fresh from that file. The fresh session almost always does better, because the wrong assumption doesn't survive the summary. Two failed attempts at the same thing is my rough rule.
Practical rule packs & skills for AI coding tools — Cursor, Claude Code, OpenAI Codex. Writing about agent workflows, project rules, and what actually ships.
That's a sharp signal: once it starts undoing its own change from two steps ago, it's arguing with itself rather than with the problem. Do you catch the undo by reading the diffs yourself, or does something else tip you off first?
I’d reset when I can no longer get the agent to state what changed between two failed attempts. The handoff I’d carry over is the failing case, evidence already checked, approaches ruled out, and the next unknown. This agent-harness exercise gives a way to test that stopping rule with a repeatable failure instead of choosing an arbitrary turn count.
does the ruled out list ever backfire tho. like an approach gets marked dead when it only failed bc of some other bug that got fixed later, and the fresh session never tries it again
Practical rule packs & skills for AI coding tools — Cursor, Claude Code, OpenAI Codex. Writing about agent workflows, project rules, and what actually ships.
It can, and your example is the case I'd worry about most. I think keeping the reason next to each entry helps: if the list says why an approach failed, a fresh session can see when that reason no longer holds and try it again.
the reason next to it is what makes it safe. id keep the commit it failed on too, so a fresh session can see if the thing that broke it has changed since
We're a place where coders share, stay up-to-date and grow their careers.
The signal that holds up: the agent proposes a fix it already reverted. Once the same wrong assumption comes back twice, more steering mostly adds text it keeps rereading. What a fresh session lacks is the list of what was ruled out and why, so that short list beats a full summary. Does /compact count as a reset for you, or as a third option?
For me /compact is a third option. It keeps the session going on a summary of the history, so a wrong assumption that made it into the summary stays. So when it loops, I'd start fresh with your short ruled-out list instead.
Fair point: the summary is where a wrong guess quietly turns into a fact. I'd keep the ruled-out list in a file the fresh session reads first, so it survives the next reset too. Do you keep one list per task, or one running file?
I build Aladdin, a job agent. My signal is the second repeat of the same fix. One retry can be a bad read. Two means the session is holding a wrong assumption, and steering just adds more words around it. When I restart, I carry over a short summary of what was learned and the verified facts, not the full transcript. The fresh worker then reads the actual page again instead of trusting the old session's picture of it. That fresh read is what breaks the loop.
One catch with carrying over the verified facts: the stuck session is usually the one that verified them, so the wrong assumption can ride along inside the summary. My guess is that's why the fresh read of the page ends up doing most of the work.
That catch is real. I keep what carries over narrow: only the facts the user gave and what was actually filled in, not the old worker's conclusions. The fresh worker still reads the form itself, and it asks the user when something is unclear instead of trusting the summary. So the summary helps with the starting point, but the fresh read does the checking.
Leaving the old worker's conclusions behind and keeping only the user's facts and what was filled in is a cleaner cut than mine.
My signal is when it starts undoing its own change from two steps ago. At that point I stop, have it write what it learned into a short notes file (what was tried, what was ruled out and why), and start fresh from that file. The fresh session almost always does better, because the wrong assumption doesn't survive the summary. Two failed attempts at the same thing is my rough rule.
That's a sharp signal: once it starts undoing its own change from two steps ago, it's arguing with itself rather than with the problem. Do you catch the undo by reading the diffs yourself, or does something else tip you off first?
I’d reset when I can no longer get the agent to state what changed between two failed attempts. The handoff I’d carry over is the failing case, evidence already checked, approaches ruled out, and the next unknown. This agent-harness exercise gives a way to test that stopping rule with a repeatable failure instead of choosing an arbitrary turn count.
does the ruled out list ever backfire tho. like an approach gets marked dead when it only failed bc of some other bug that got fixed later, and the fresh session never tries it again
It can, and your example is the case I'd worry about most. I think keeping the reason next to each entry helps: if the list says why an approach failed, a fresh session can see when that reason no longer holds and try it again.
the reason next to it is what makes it safe. id keep the commit it failed on too, so a fresh session can see if the thing that broke it has changed since