Hora v2.2 won the test and was promoted. A week later, logistics changed the delivery day for one route and the agent's table went stale. The conversations handled by Juliana stayed low, but missed delivery-time promises went over the limit. Promoting and rolling back are planned steps: in this lesson you write both before you need them.
1What comes with a promotion: phases, watch, and a trigger to go back
2Calculate the trigger so you don't roll back by chance
3Roll back fast, warning first and keeping the version
4The version history and the cycle's own authority limit
Goal: treat promotion and rollback as planned steps: promote with a written way back, and roll back by the trigger, with no hunt for someone to blame.
01 · From promotion to the next candidate
Promoting is not the end of the cycle. Click each step.
Kept or rolled back, the lesson goes into the history and feeds the next hypothesis
Click each block in the diagram to see what it does.
AI Management Area, rule 3: promoting and rolling back are steps, not accidents. Horizonte numbers: fictional, for study, in relative weeks.
02 · Four rules for promoting and going back
G2 put the process under version control and G3 promoted autonomy actions. Here the version is that of the agent package (prompt, context, model, tools) that came out of a LOOP-R cycle.
1. The promotion carries the rollback plan
Carlos approved v2.2 in week 7 with the rollback plan already written: a watch window of two weeks; a trigger of missed delivery-date promises above 3.4% in a week, or one critical error; who decides, Patrícia; how to go back, Rafael points Hora to v2.1 within 2 hours.
Deciding the rollback after the problem hits is deciding in a panic, and it almost always turns into an argument about blame.
RulePromoting needs approval from someone who did not propose it. Rolling back by the written trigger is the duty of whoever operates it, and it does not wait for a meeting.
2. The trigger allows for chance
Missed promises are a rare event: across 160 delivery-date conversations a week, the rate moves on its own. The trigger adds two standard errors for the week to the pilot rate (1.5%): 2 × √(0.015 × 0.985 ÷ 160) ≈ 1.9 points. Limit: 3.4%.
So a week with 4 missed promises (2.5%) does not roll back; a week with 7 (4.4%) does. It is the same calculation LOOP-R uses for tolerance.
CounterexampleIn the LOOP-R example, 6 complaints against 1, out of 300 per version, with zero tolerance, threw out a good idea by chance.
3. Roll back fast, with no one to blame
In week 8, logistics changed the delivery day for the Sorocaba route and the table was not updated. Missed promises: 7 of 158 (4.4%), above 3.4%. The conversations taken over by Juliana stayed at 19.0%: the target was doing fine, the guardrail fell.
Patrícia decided by the trigger; Juliana was told beforehand; Rafael pointed back to v2.1 within 2 hours; v2.2 was kept in storage. Carlos was informed with the number.
A cause, not someone to blameThe table had no expiry date and no notice for route changes. The fix is in the document (owner, expiry, notice), not in Valdir.
4. Version history and the cycle's authority limit
Each version gets one line in the history: what changed, why, who approved it, the status. The version that was rolled back is not erased; it becomes a memory of what did not work and why.
The cycle itself has an authority limit. In LOOP-R, L0 only proposes; L1 promotes with one person's approval for each change; L2 promotes on its own, and only after at least five promotions at L1 with no rollbacks. Horizonte is at L1.
Next stepv2.3: the same table, with Valdir as owner, a weekly expiry date, and a notice for route changes. It goes back through the 30 cases and a new pilot.
03 · Templates to copy
The promotion log, the rollback log, and the version history. The tab next to this one shows Horizonte's.
Promotion log
PROMOTION · agent <name> · from v<current> to v<candidate> · LOOP-R cycle no. <...>
Proposed by: <name> · Approved by (not the proposer): <name> on <date>
Decision card attached: [ ] yes
PHASES: <a slice first, then everyone, or everyone at once, and why>
WATCH WINDOW: <n weeks> · guardrails checked <how often>, by <name>
ROLLBACK TRIGGER (written now)
<guardrail> above <limit> in <period>
limit calculation: <pilot rate> + 2 × sqrt(p × (1 - p) ÷ <N for the period>) = <...>
or: 1 critical error
WHO DECIDES THE ROLLBACK: <name> · without waiting for approval
HOW TO GO BACK: <steps, who does each one, in how long>
WHO IS TOLD BEFOREHAND: <names> · PREVIOUS VERSION KEPT AT: <where>
Rollback log
ROLLBACK · agent <name> · from v<promoted> to v<previous> · <date and time>
Trigger that fired: <indicator> = <value> (<n> of <N>) · written limit: <...>
Target of the change in the same week: <value> (for the log, not for deciding)
Decided by: <name> · notified beforehand: <names> · executed by: <name> in <time>
Rolled-back version kept at: <where>
LIKELY CAUSE: <what changed in the world or in the agent>
PROPOSED FIX: <...> · new candidate: v<...> · back to the 30 cases and the pilot: <dates>
LESSON FOR THE HISTORY: <one sentence>
Agent version history
HISTORY · agent <name> · owner: <name> · cycle authority limit: [ ] L0 proposes [ ] L1 a person approves each change [ ] L2 promotes on its own
| Version | What changed | Why (evidence) | Approved by | Status | Date |
|---|---|---|---|---|---|
| <v> | <...> | <...> | <name> | in use / kept / rolled back / proposed | <...> |
Rule: no row is ever deleted. A rolled-back version stays with the reason.
L2 only after 5 promotions at L1 with no rollback.
Prompt · review the rollback plan
You are a change reviewer. Below is the promotion record for a new
version of an AI agent. Do not decide whether the promotion is good. Point out:
1. Whether the rollback trigger has a number, a period and the math behind the limit.
2. Whether the limit is so tight that normal variation sets it off, or so
wide that it never fires.
3. Whether who decides the rollback, the steps and the time are written down.
4. Who would be affected by the rollback and is not on the notified list.
Record:
"""<paste>"""
Simulation: illustrative example
Hora v2.1 → v2.2 → v2.1 · Horizonte, fictional study company · relative weeks
Week 7, promotion: proposed by Patrícia, approved by Carlos with the decision card from cycle 0001. All conversations on v2.2. Watch window of 2 weeks.
Trigger: missed delivery-date promises above 3.4% in a week (1.5% from the pilot + 1.9 points), or 1 critical error. Decides: Patrícia. Rollback: Rafael points it back to v2.1 within 2 h. Notified beforehand: Juliana.
Week 8, rollback: 7 of 158 promises missed (4.4%). Taken over by Juliana: 30 of 158 (19.0%). Rolled back in 2 h, v2.2 kept, Carlos informed.
Likely cause: the Sorocaba route changed days and the table had no expiry date or change alert.
Next candidate: v2.3, route table with Valdir as owner, weekly expiry and a change alert.
Version
What changed
Status
2.0
G3 dossier: role, constitution, matrix
kept
2.1
Lock: the day's total for the same customer above R$ 5 mil goes to approval (case T-27)
in use
2.2
Route table in the context package (cycle 0001)
rolled back in week 8: missed promises 4.4% > 3.4%
2.3
Route table with an owner, weekly expiry and a change alert
proposed
Cycle authority limit: L1. No row deleted.
Promoting without a trigger. When the problem hits, nobody knows whether to roll back.
A trigger with no math. Too tight, and it rolls back by chance; too wide, and it never fires.
Waiting for a meeting to roll back. Rolling back by the written trigger needs no approval.
Deleting the version that failed. It is the memory of what did not work.
Looking for someone to blame. The question is what changed and what protection was missing.
04 · Simulation: Carlos doesn't want to roll back
A simulated conversation for practice. The trigger fired and the owner liked the new version. Answer with what is written in the promotion record.
Carlos · owner of Horizonte (simulation)Horizonte Distribuidora, fictional study company
Choose your answer. The comment appears right after, and the conversation continues based on your choice.
05 · Practice
Horizonte's logs are for practice; the promotion and the rehearsed rollback of your agent are for real.
1 · Hora's promotion, rollback, and history
30 min
Using the numbers from this lesson, fill in the promotion log for v2.2 (week 7) and the rollback log for week 8. Redo the trigger calculation. Build Hora's history with versions 2.0 to 2.3 and mark the cycle's authority limit.
Done when: the promotion has an approver different from whoever proposed it, a watch window, a trigger with a number and the calculation (1.5% + 1.9 = 3.4%), who decides the rollback, the steps with times, and the people notified; the rollback has the number that fired it (7 of 158), the probable cause, and the next candidate; and the history has 4 lines, none erased.
2 · Your agent's rehearsed rollback
45 min
For the change in your lesson 5 cycle (or for your agent's next change), write the promotion log with the calculated trigger. Then rehearse the rollback in a test environment or outside business hours: point to the previous version, time it, and check that nothing was lost.
Done when: the promotion log is saved with the trigger as a number and the calculation; the rollback was rehearsed once with the time measured; your agent's version history exists; and whoever decides the rollback confirmed in writing that they know this.
3 · The conversation with Carlos
10 min
Run the simulation above until you reach the top score. If you get one wrong, read the comment and start over.
Done when: the simulation ended on "Roll back without drama, improve with method" with 8 of 8 points.
06 · Quick test
4 questions. The explanation appears after your attempt.
Question 1 of 4
After the promotion, v2.2's goal was still great, but a guardrail went past the written limit. Which one counts?
Correct
A guardrail exists so that a nice-looking goal does not hide damage. If it drops past the limit, the version rolls back.
Who needs to approve a rollback done by the written trigger?
Correct
Promoting requires approval from someone who did not propose it; going back to the known version, by the agreed trigger, does not wait for a meeting. Carlos is informed with the number.
Not this one
Promoting requires approval from someone who did not propose it; going back to the known version, by the agreed trigger, does not wait for a meeting. Carlos is informed with the number.Review: Four rules for promoting and going back
Question 4 of 4
From when can a LOOP-R cycle promote on its own (L2)?
Correct
The cycle itself earns authority limit like an agent: with evidence and one level at a time. Horizonte just rolled back a promotion, so it stays at L1.
Not this one
The cycle itself earns authority limit like an agent: with evidence and one level at a time. Horizonte just rolled back a promotion, so it stays at L1.Review: Four rules for promoting and going back
Quiz complete
You got {a} of {n} right. Get them all right and the lesson is marked as complete.
Choose an answer. The explanation appears after you try.
07 · Summary
Promotion with a way backWatch window, trigger as a number, who decides, and how to roll back, written beforehand.
Trigger with a numberThe pilot rate plus two standard errors for the week: it neither rolls back by chance nor never rolls back.
Roll back with no one to blameKnown version, warning first, nothing erased; then the cause and a safeguard.
History and authority limitEvery version is logged; the cycle moves from L1 to L2 only with evidence.
