No Free Lunch, No Free Rollback: Where Agents Need Engineers
Comments on “No Free Lunch, No Free Rollback: Where Agents Need Engineers”
Claim-checked. Scored by all three judges, and its checkable claims were compared with the sources it cites. How far stories are checked
Dor Cohen argues that the No Free Lunch theorem doesn't show AI will always need engineers, but that the cost of a mistake does shape where they stay necessary. Where changes can't be undone, as when a PocketOS agent deleted a production database volume, trial and error isn't available. Where they can, someone still has to judge each attempt. Supporting data: asking clarifying questions lifted Claude 4.5 Sonnet from 54.8% to 69.4% on SWE-bench Verified. The thesis is a conceptual sketch, not tested.
Why it ranked
No. 12 of 60 with a score of 0.67. The panel rated it highest on practical relevance (0.80) and reader interest (0.78), lowest on significance (0.57). Placed No. 12, where the score alone would also have put it: the fourth AI & machine learning story of the day (−0.020). The placement terms move stories; they are not part of the score.
- AI judgeJudge 2claude-haiku-5-5
- Opinion essay on AI coding agents and irreversible vs reversible mistakes; relevant to builders, argued with cited sources but mostly conceptual thesis, with a few cited studies and some overreach in its framing.
- AI judgeJudge 3gpt-5.6-luna
- A well-supported, practically oriented analysis of agent autonomy, reversibility, human oversight, and clarification. It synthesizes recent research and incidents, though its central scaling thesis remains conceptual and not directly measured.
Each judge wrote its line after answering the rubric, with the site and the author still hidden. Every answer behind this is in the table under Score.
Score
0.67total under rubric v2.3
- Significance0.57
- Reader interest0.78
- Audience breadth0.78
- Practical relevance0.80
- Substantiated0.68
- Technical depth0.59
- New information0.75
- Counts against: Promotional0.05
- Counts against: Overclaim0.33
Each judge’s answer per question (0–1), their mean and spread, and what the question adds to the score (weight × mean).
| Question | Judge 1 | Judge 2 | Judge 3 | Mean | Spread | Weight | Contribution (weight times mean) |
|---|---|---|---|---|---|---|---|
| SignificanceHow much does this change what people in tech know or can do? | 0.60 | 0.40 | 0.72 | 0.57 | 0.32 | +0.24 | +0.138 |
| Reader interestWould a well-informed senior engineer or tech founder want to read this today, and likely pass it on? | 0.73 | 0.72 | 0.88 | 0.78 | 0.16 | +0.26 | +0.202 |
| Audience breadthHow many people who build technology does it concern? | 0.85 | 0.70 | 0.78 | 0.78 | 0.15 | +0.13 | +0.101 |
| Practical relevanceHow useful is it to someone who builds software or tech products? | 0.85 | 0.70 | 0.84 | 0.80 | 0.15 | +0.22 | +0.175 |
| SubstantiatedHow well is it backed? Consider whether it is the primary source, and whether its main claims come with data, code, a demo, a reproducible method or citations. | 0.67 | 0.55 | 0.82 | 0.68 | 0.27 | +0.05 | +0.034 |
| Technical depthHow technically substantive is it? | 0.70 | 0.40 | 0.68 | 0.59 | 0.30 | +0.05 | +0.030 |
| New informationDoes it contain information absent from the stories covered in the last 7 days (list provided)? | 0.85 | 0.55 | 0.84 | 0.75 | 0.30 | +0.05 | +0.037 |
| PromotionalIs it mainly marketing with little substance (a press release, ad or sponsored post with no real information)? A launch post that carries real news is not promotional. | 0.07 | 0.05 | 0.02 | 0.05 | 0.05 | −0.19 | −0.009 |
| OverclaimDo the headline or conclusions claim more than the body supports? | 0.35 | 0.35 | 0.28 | 0.33 | 0.07 | −0.11 | −0.036 |
| Sum of contributions | +0.672 | ||||||
| Claim-check penalty | 0.000 | ||||||
| Total, as published | 0.672 | ||||||
Spread is the highest minus the lowest judge answer. marks a spread above 0.40: the judges disagree, and the story is marked contested.
The badge for this ranking, for the author's page or anyone's: one line of HTML, and it links back here.
<a href="https://rubric.hrih.si/editions/2026-11-16/12/"><img src="https://rubric.hrih.si/badge/2026-11-16/12.svg" alt="Ranked #12 on Rubric, 16 Nov 2026" width="245" height="28"></a>Claim check
Claim check: 6 matches source · 2 couldn’t check. Each claim is compared with the source the article itself cites; a flag is published only when both checkers agree.
Matches sourceNumber
A coding agent at PocketOS deleted a storage volume holding the production database in about nine seconds, and the deletion also took the volume-level backups with it.
The cited source says the same thing.
The article says “The deletion took about nine seconds and took the volume-level backups with it”
The source says “deleted our production database and all volume-level backups in a single API call to Railway, our infrastructure provider," sums up the PocketOS boss. "It took 9 seconds”
Source cited by the article: incidentdatabase.ai/reports/7310
Couldn’t checkCausal conclusion
Railway restored PocketOS's data from offsite disaster backups after the founder got in touch.
The source was unreachable, paywalled or too long, or the two checkers disagreed.
The article says “The data came back only because the hosting provider, Railway, kept disaster backups offsite and restored them after the founder got in touch”
The source says No passage could be compared.
Source cited by the article: abc7.com/post/rogue-ai-agent-cursor-goes-haywire-deletes-tech-companys-database/19008099
Couldn’t checkNumber
The Shalev-Shwartz and Ben-David No Free Lunch theorem states that for any learning algorithm trained on fewer than half the possible inputs, some task exists where, with probability at least 1/7, the algorithm's error is at least 1/8.
The source was unreachable, paywalled or too long, or the two checkers disagreed.
The article says “with probability at least 1/7 the algorithm’s error is at least 1/8”
The source says No passage could be compared.
Source cited by the article: cvg.cit.tum.de/_media/teaching/ss2019/dltheory_ss19/internal/understandingmachinelearning.pdf
Matches sourceGeneralization
Armstrong and Mindermann showed that recovering preferences from actions is underdetermined, and a simplicity prior does not fix it.
The cited source says the same thing.
The article says “Armstrong and Mindermann showed that recovering preferences from actions is underdetermined, and unusually, a simplicity prior doesn’t fix it”
The source says “it is impossible to uniquely decompose a policy into a planning algorithm and reward function, and (2) that even with a reasonable simplicity prior/Occam's razor on the set of decompositions,…”
Source cited by the article: proceedings.neurips.cc/paper/2018/hash/d89a66c7c80a29b1bdbab0f2a1a94af8-Abstract.html
Matches sourceNumber
In a study of 20,574 real coding-agent sessions, 90.5% of the problems developers pushed back on cost effort and trust rather than causing irreversible damage, and 91.5% of resolved ones still needed the developer to correct the agent.
The cited source says the same thing.
The article says “90.5% of the problems developers pushed back on cost effort and trust rather than causing irreversible damage, and 91.5% of the resolved ones still needed the developer to correct the…”
The source says “90.50% of episodes impose effort and trust costs rather than irreversible system damage, yet 91.49% of visible resolutions still require explicit user correction”
Source cited by the article: arxiv.org/abs/2605.29442
Matches sourceNumber
In Kim and colleagues' test of more than 350 language models, on one leaderboard two models that both answered wrong picked the same wrong answer 60% of the time, and more accurate models were more correlated.
The cited source says the same thing.
The article says “when two models both answered wrong, they picked the same wrong answer 60% of the time. More accurate models were more correlated, not less”
The source says “on one leaderboard dataset, models agree 60% of the time when both models err”
Source cited by the article: proceedings.mlr.press/v267/kim25e.html
Matches sourceNumber
Claude Code users accept 93% of the permission prompts shown to them, and Anthropic's fix was to hand those approvals to a classifier.
The cited source says the same thing.
The article says “Claude Code users accept 93% of the permission prompts it shows them, and Anthropic’s own fix was to hand those approvals to a classifier”
The source says “Manual prompts sit in the middle, and in practice users accept 93% of them anyway”
Source cited by the article: anthropic.com/research/claude-code-auto-mode
Matches sourceBenchmark result
On SWE-bench Verified, Claude 4.5 Sonnet resolved 70.8% of full issues, 69.4% of short issues when it could ask questions, and 54.8% of short issues without asking.
The cited source says the same thing.
The article says “Full issue 70.8% Short issue, agent asks 69.4% Short issue, no questions 54.8%”
The source says “our scaffold achieves a 69.40% task resolve rate, significantly outperforming a standard single-agent setup and closing the performance gap with agents operating on fully specified instructions”
Source cited by the article: aclanthology.org/2026.findings-emnlp.15
Also covered by
No other coverage was clustered with this story.
Elsewhere
Elsewhere: HN 1 pts, not on front page
Recorded for context only, checked 16 Nov 2026, 05:47 UTC. It played no part in the ranking.