Quirin Schlegel

Loading

Supervision was a schedule — a monitor that varies when it looks beats one that looks often, and omitting oversight became a crime in Italy twelve days ago

Inspection of an agent's trajectory is a schedule, not a checkbox: adapting when you look beats shortening the interval, and Italy already attaches criminal and discoverable-record consequences to an oversight omission that the AI Act's own Article 14 will not reach until December 2027.

12 October 202616 verified claims7 sources

A reviewer opens the pull request at the end of a long agent run. That is the first time anyone looked, and for most teams the only inspection point.

A paper posted to arXiv on 8 October makes that moment the variable. Researchers at Singapore Management University and University College London built a runtime monitor that varies when it inspects an agent's trajectory instead of inspecting on a fixed clock.

Flat:

1. Across 300 SWE-bench Lite tasks: 42.33% resolved on mini-swe-agent against 17.00% unmonitored, and 39.67% on Moatless against 24.00%. The abstract reports these as "+25.33%" and "+15.67%" — percentage points, not relative gains; relative, the first is +149%.

2. The mechanism is not more supervision. Fixed-interval monitors at every 3, 5 and 7 steps landed 33.33%, 34.00% and 26.33% — not monotone in the interval. The lever is tightening inspection after a severe correction and relaxing it after a mild one. It also cost less: 1.03M tokens per task against 1.42M unmonitored.

3. Before you read this as settled: on Moatless the strongest baseline reached 35.33% against Cadence's 39.67%, a gap the paper reports as not statistically significant. One agent shows a result; the other a direction.

4. Now the European part, and it is not Article 14. Article 14's human-oversight duty binds Annex III high-risk systems from 2 December 2027 and Annex I from 2 August 2028 after the Digital Omnibus. What already happened is Italian. D.Lgs. 160/2026 has been in force since 30 September: omitting human-oversight measures on a high-risk system is a criminal offence, with corporate liability attached, and in a civil claim the court can order production of your records while causation is presumed unless you rebut it. D.Lgs. 179/2026 enters into force on 23 October and names who asks: ACN, the cyber agency. In Germany that seat is the Bundesnetzagentur.

Your coding agent is not a high-risk system. The system it helps you build may be, and its trace is then discoverable evidence about how you built it.

My Monday: take ten agent runs from last week. For each, note the step where it went wrong and the step where anyone looked. If the only match is the pull request, your interval is one.

An LLM watching a trajectory is not human oversight — and a monitor you cannot timestamp is not oversight.

When did you last look mid-run?

#AgenticAI #EUAIAct #CodingAgents #AIGovernance #SoftwareEngineering

AI disclosure: the narration voice, the cover art and the brand ident animation in this video are AI-generated. The script, the claims and the source checks are mine.

Evidence

What this rests on

Every factual claim in the edition above, stated on its own and checked before publication. If one of these is wrong, the argument is wrong — so they are listed where you can check them.

Cadence, a runtime monitor that varies its inspection schedule, resolves 42.33% of 300 SWE-bench Lite tasks on mini-swe-agent against 17.00% for the same agent with no monitor.
Why it mattersThe comparator is the same agent unmonitored, so the figure isolates the monitor rather than the model. It is the number to quote when arguing for an inspection budget.
arxiv2026cadence
On Moatless the same monitor resolves 39.67% against 24.00% unmonitored, a 15.67 percentage-point gap.
Why it mattersTwo agents, two magnitudes. A practitioner should expect the benefit to depend on the harness, not to be a constant.
arxiv2026cadence
The abstract's "+25.33%" and "+15.67%" are percentage-point differences in resolve rate, not relative improvements; relative, the first is approximately +149%.
Why it mattersQuoting the abstract's figure as a relative gain in a review would misstate the result by roughly sixfold in the monitor's favour.
arxiv2026cadence
Fixed-interval monitoring on mini-swe-agent is not monotone in the interval: every 5 steps gives 34.00%, every 3 gives 33.33% and every 7 gives 26.33%.
Why it mattersIt rules out the obvious intervention. Inspecting more often is not the lever; adapting when you inspect is.
arxiv2026cadence
On mini-swe-agent the monitored run cost 1.03M tokens per task against 1.42M unmonitored, while on Moatless it cost 0.48M against 0.32M.
Why it mattersSupervision paid for itself on the weaker agent and cost money on the stronger one, which is what decides where an inspection budget goes.
arxiv2026cadence
On Moatless the strongest baseline monitor reaches 35.33% against Cadence's 39.67%, and the paper reports that gap as not statistically significant.
Why it mattersIt is the limit of the result. One agent shows a win; the other shows a direction, and an honest summary says so.
arxiv2026cadence
Italy's D.Lgs. 9 settembre 2026, n. 160, in force since 30 September 2026, makes the omission of human-oversight measures for a high-risk AI system a criminal offence under a new Art. 437-bis of the criminal code.
Why it mattersIt is the first EU member state to attach personal criminal liability to an oversight omission, and it is already in force rather than pending.
normattiva2026reati
The same decree adds Art. 25-vicies to D.Lgs. 231/2001, making AI-committed offences predicate offences for corporate liability, with disqualification sanctions that include a ban on advertising.
Why it mattersCorporate liability is what turns an engineering omission into a board-level exposure, and the advertising ban is a commercial rather than financial sanction.
normattiva2026reati
Under D.Lgs. 160/2026 a court may order a provider or user to produce records and technical documentation about the algorithm, may draw adverse inferences from an unjustified refusal, and applies a rebuttable presumption of the causal link in the victim's favour.
Why it mattersIt converts machine-generated traces into discoverable evidence and shifts the burden, so the absence of a record is itself a litigation position.
normattiva2026reati
D.Lgs. 7 ottobre 2026, n. 179 was published in Gazzetta Ufficiale Serie Generale n. 234 of 8 October 2026 and enters into force on 23 October 2026.
Why it mattersIt is the live deadline: eleven days from this edition, and the instrument that settles which Italian authority asks the questions.
normattiva2026adeguamen
That decree designates AgID and ACN as Italy's national AI authorities, AgID as notifying authority and ACN, the national cybersecurity agency, among the market surveillance authorities.
Why it mattersThe authority that would eventually ask about a high-risk deployment is a cybersecurity agency in Italy and a telecoms regulator in Germany, which changes what it asks.
normattiva2026adeguamen · bundesamtfu2026kimig
Under D.Lgs. 179/2026 Art. 41, decisions on hiring, dismissal, performance evaluation and discipline may not rest solely on automated processing, and the final decision is reserved to a natural person exercising effective and autonomous authority.
Why it mattersAnyone shipping agentic tooling into an HR workflow in Italy has a hard constraint from 23 October, independent of the AI Act's high-risk timetable.
normattiva2026adeguamen
AI Act Article 14 requires high-risk systems to be designed so they can be effectively overseen by natural persons while in use, and binds Annex III systems from 2 December 2027 and Annex I systems from 2 August 2028.
Why it mattersIt is the article a practitioner reaches for and the wrong one for today: the duty is fourteen months out while Italy's criminal limb is already in force.
euaiactexpl2026human · eurlex2026regulatio
Article 14's oversight duty runs to natural persons, so an LLM runtime monitor inspecting an agent trajectory does not on its own discharge it.
Why it mattersIt stops a monitor being presented as compliance. The monitor can be part of the interface tooling that makes oversight effective; it cannot be the overseer.
euaiactexpl2026human
The EUR 35,000,000 or 7% ceiling quoted in coverage of the Italian decrees is AI Act Article 99(3), for the Article 5 prohibitions; operator duties and Article 50 transparency sit at Article 99(4), EUR 15,000,000 or 3%.
Why it mattersThe headline figure is for prohibited practices, not for the duties a practitioner is likely to breach, and quoting it for the latter overstates the exposure by more than half.
euaiactexpl2026penalties
Regulation (EU) 2026/1744, the Digital Omnibus on AI, in force 27 July 2026, moved Annex III high-risk to 2 December 2027, Annex I to 2 August 2028 and member-state sandboxes to 2 August 2027.
Why it mattersIt is the instrument that makes every pre-August-2026 compliance calendar wrong, including most of the commentary a practitioner will find by searching.
eurlex2026regulatio

Sources

Every source for this edition, as one PDF BibTeX — all editions

Corrections

What changed after publication

A coverage date is not an event date, and this time the check changed the edition. An English-language trade piece headlined "Italy's AI decree 179/2026 takes effect October 23" and then dated the instrument to 9 September 2026, published 15 September in Gazzetta Ufficiale n. 214, while attributing to it the criminal offence, the 231/2001 predicate and the civil-liability provisions. Those are two different decrees. The Normattiva register shows D.Lgs. 179/2026 dated 7 October 2026, published in GU Serie Generale n. 234 of 8 October 2026, in force 23 October 2026, and titled for the national authorities' powers and AI in training. The September dates and every liability provision belong to D.Lgs. 9 settembre 2026, n. 160, GU n. 214 of 15 September 2026, in force 30 September 2026. Had the conflation stood, the edition would have told readers a criminal liability regime arrives on 23 October when it has been live since 30 September, and would have cited the wrong instrument for all of it. Both decrees are cited separately and each date comes from the register record, not the coverage.

Affiliations taken from the work's own front matter, and available this time. The arXiv abs record for 2610.12269 prints no institutions, so the HTML rendering at arxiv.org/html/2610.12269v1 was tried next and carries them: Minxing Wang and Yintong Huo at Singapore Management University, He Ye and Earl T. Barr at University College London. The post names both institutions on that basis. No institution was taken from an aggregator.

Identifier verified before any figure was used, not the topic. 2610.12269's title matches the story exactly and its YYMM is consistent with an 8 October 2026 submission. Four adjacent October papers returned by the same searches were considered and not used: 2610.10610 (Code Understanding is a Bottleneck for Coding Agents), 2610.09944 (AgentTime), 2610.09264 (Package Hallucination Attacks through rule files, which overlaps the 9 October edition), and 2610.13361 (AgentSWE).

A conflation smell inspected and found real but mislabelled. The abstract's "+25.33%" and "+15.67%" match the headline gaps exactly, which is the smell. Reading the tables, they are percentage-point differences in resolve rate (42.33 - 17.00 and 39.67 - 24.00), not relative gains. Published as percentage points with the relative figure given once, so a reader cannot take the abstract's wording for a 149% improvement.

A metric dropped rather than published, because two sources disagree. The prison-term ranges in Art. 437-bis are reported as 1-5 years for danger to life or safety and 2-8 where State security is engaged by one English-language source, and as 2-8 years for concrete demonstrable danger, 2-6 for alteration by an outsider and 3-10 in national-security cases by the Italian legal press. No term appears in the post or the video; the edition states that the omission is a criminal offence and leaves the sentencing out.

A limit stated rather than smoothed over. On Moatless the strongest baseline reaches 35.33% against Cadence's 39.67% and the paper calls that gap not statistically significant. The post carries it as item three rather than leaving the reader with two clean wins.

A ceiling corrected against the Article it actually sits in. Coverage of the Italian decrees quotes EUR 35,000,000 or 7%. That is Article 99(3), for the Article 5 prohibitions. Operator duties and Article 50 transparency are Article 99(4): EUR 15,000,000 or 3%. The first comment says which is which rather than repeating the larger number.

Cite keys reconciled against the published vault before the archive ran, not after. Four keys were regenerated for documents the vault already holds - euaiactexpl2026human, euaiactexpl2026penalties, eurlex2026regulatio and bundesamtfu2026kimig - by matching publisher and title strings to the earlier spelling, and three were created after checking all 390 existing keys: arxiv2026cadence, normattiva2026reati and normattiva2026adeguamen. The two Italian decrees were titled from their subject rather than their number, because citekey() derives its tail from the title's first token and "D.Lgs." in both would have collided on "dlgs".