Interrupting an agent run mid-loop: when to stop research, redirect the writer, or kill the draft
A supervised agent pipeline gives you three real levers over an in-flight post: what happens between stages, what you change on the next stage's brief, and whether the run continues at all. The interesting question is not whether you can hit a giant red button. It is which lever to reach for when the log or the draft tells you something is off.
This is a working guide to those decisions. It assumes you are running the RankMill loop: a research agent, a writer, an editor, and a human approval checkpoint between them. It focuses on the signals you can actually read in the live agent log and the draft, and what each signal is telling you to do.
The three intervention points, in order
There are three places a human can act on a run without waiting for the finished post:
- After research, before drafting. The research agent has proposed a topic, an angle, keywords, and a working outline. You either approve, redirect, or kill.
- After drafting, before publish. The writer has produced a full draft. The editor agent is your redirect tool here, taking natural-language change requests instead of hand edits. Kill is still on the table if the whole angle is wrong.
- The publish gate itself. Nothing ships without your say-so. In practice this rarely fires as a hard stop, because the previous two checkpoints should have caught anything worth catching.
The mistake most people make is treating the first checkpoint as a formality and the second one as the place to fix everything. It is much cheaper to redirect research than to redirect the writer, and much cheaper to redirect the writer through the editor agent than to rewrite by hand.
Reading the research log: signals that mean act
The research agent narrates its reasoning as it works. That log is the cheapest place to catch a bad run, because you are stopping before anyone spent effort drafting.
Watch for these signals:
The angle drifted. You asked for a post on onboarding, and the log is now spelunking through pricing pages. That is a signal to redirect, not to let it play out and hope the outline snaps back. Tell it what the piece is actually for.
The competitor set is wrong. If the agent is anchoring on sites that do not match your positioning, the resulting outline will inherit their frame. Cheap in-flight fix: correct the competitor list at the next checkpoint and let it re-scope.
Search intent got misread. A "how to" query treated as commercial, or the reverse, will produce a draft that misses the reader entirely. If the log's intent classification looks off, redirect before drafting starts.
The outline is generic. If the proposed structure could belong to any brand in the category, you are about to get a generic post. Kill or redirect with a sharper angle. Approving a generic outline and hoping the writer saves it is how thin content gets shipped.
The keyword list is stuffed or unrelated. Long tails that nobody in your audience would type, or a pile of head terms with no through-line, mean the SEO framing is off. Fix it here. This is the last place fixing it is free.
If the log looks clean and the outline reflects your positioning, approve and move on. Do not redirect for taste when the substance is right.
Between drafting and publish: the redirect versus kill decision
Once the writer has produced a draft, you have two useful moves and one expensive one. The useful moves are approving the draft or sending change requests to the editor agent. The expensive one is killing the run and starting over.
The editor agent is the right tool when the bones are good and the surface needs work. Good editor requests look like this:
- "Cut the intro to two sentences and get to the checklist faster."
- "Replace the second example with something a solo founder would recognize."
- "Tighten the middle section, it repeats the point in the first paragraph."
- "Swap the closing CTA for something softer, this reader is not ready to buy."
Notice what these have in common. They act on structure, emphasis, and voice, not on the underlying claim. Editor rounds are cheap. Use them.
Kill the run when one of these is true:
The angle is wrong. No amount of editing turns a post about the wrong audience into a post about the right one. The whole frame has to change, which means new research.
The facts are wrong. If the draft is repeating claims that are not in your source material and not on your site, do not chase them down with hand edits. Kill it, fix the input, and rerun. Every fact you patch in by hand is a fact the next run will get wrong the same way.
Voice is off in a way edits will not fix. If the writer is producing marketing-brochure prose and your brand voice is practitioner-to-practitioner, the fix is upstream in writing rules, banned words, and brand voice configuration. Edit the config, then rerun. Patching one draft leaves the same trap set for the next one.
You do not actually want this post. Reading the draft, you realize the topic is not worth the post budget it will cost. Kill it. Post-based pricing is generous with this decision, because you are not paying for an idle seat either way.
The general rule: if you are writing an editor request that starts with "actually, the whole point of this piece is," you are past the editor's job. Kill and restart with a better brief.
The case for killing instead of editing your way out
There is a specific failure mode to watch for: the death-by-editor spiral. You send one change request, then another, then a third, and by the fifth round you have a Frankenstein draft that is worse than the first one. Every round changed something, and nothing is coherent anymore.
The signal that you are in this spiral is simple. Are your requests getting smaller and more surgical, or are they contradicting each other? Small and surgical means the piece is converging. Contradicting means you do not actually know what you want, and the editor cannot solve that.
Kill the run. Write a one-paragraph brief for yourself. Rerun. You will finish faster and the post will be more coherent.
The same holds for hand editing. If you have spent thirty minutes rewriting the intro and the middle no longer connects, stop. The problem is not the intro. The problem is that the brief was ambiguous and the writer picked a different resolution than you wanted. Fix the brief.
A short decision tree
At the research checkpoint:
- Angle right, outline specific, keywords sensible: approve
- Angle drifted, competitors off, or intent misread: redirect
- Outline is generic or the topic is not worth the post: kill
At the drafting checkpoint:
- Bones good, surface needs work: editor round
- Repeated editor rounds, contradicting each other: kill
- Wrong angle, wrong facts, or voice broken at the config level: kill and fix upstream
- Draft is right: approve and publish
Two habits make this tree cheap to run. Read the research log while it is streaming, not after, so you arrive at the checkpoint knowing what you will say. And when you write an editor request, ask yourself whether it fixes the piece or hides a deeper problem. If it hides, kill instead.
Control is not a button you press when things go wrong. It is a set of small decisions made at the right moments, in the right order. The loop is designed to make each of those moments cheap. Use them.