Anatomy of an Agent Harness: 30 decisions, invariants and corner cases
Comments on “Anatomy of an Agent Harness: 30 decisions, invariants and corner cases”
Claim-checked. Scored by all three judges, and its checkable claims were compared with the sources it cites. How far stories are checked
The paper, "Anatomy of an Agent Harness: 30 decisions, invariants and corner cases," catalogues the decisions a harness must make around a language model: building requests, running tools, permissions, transcripts, failure recovery and display.
Why it ranked
No. 22 of 142 with a score of 0.69. The panel rated it highest on technical depth (0.94) and new information (0.91), lowest on significance (0.54). Placed No. 22, No. 15 by score alone: the sixth AI & machine learning story of the day (−0.060), and close to “Terminal-Bench 4.1: 23 of 66 tasks have scoring or data problems”, No. 5 on Wednesday 11 November 2026 (−0.003). The placement terms move stories; they are not part of the score.
- AI judgeJudge 2claude-haiku-5-5
- Detailed, substantive design-catalogue for AI agent harnesses with explicit test methods and evidence tags; very practical for builders, but self-referential, single-author, and narrow in audience and significance.
- AI judgeJudge 3gpt-5.6-luna
- A technically deep, unusually systematic paper turns agent-harness design into testable invariants, with explicit evidence tags, corner cases, and limitations. Its self-assessment and incomplete evidence temper significance, but it is highly actionable for agent builders.
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
0.69total under rubric v2.3
- Significance0.54
- Reader interest0.77
- Audience breadth0.57
- Practical relevance0.84
- Substantiated0.72
- Technical depth0.94
- New information0.91
- Counts against: Promotional0.04
- Counts against: Overclaim0.16
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? | 0.62 | 0.35 | 0.65 | 0.54 | 0.30 | +0.24 | +0.130 |
| Reader interestWould a well-informed senior engineer or tech founder want to read this today, and likely pass it on? | 0.74 | 0.70 | 0.88 | 0.77 | 0.18 | +0.26 | +0.201 |
| Audience breadthHow many people who build technology does it concern? | 0.59 | 0.40 | 0.72 | 0.57 | 0.32 | +0.13 | +0.074 |
| Practical relevanceHow useful is it to someone who builds software or tech products? | 0.90 | 0.80 | 0.82 | 0.84 | 0.10 | +0.22 | +0.185 |
| 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. | 0.78 | 0.60 | 0.78 | 0.72 | 0.18 | +0.05 | +0.036 |
| Technical depthHow technically substantive is it? | 0.98 | 0.90 | 0.94 | 0.94 | 0.09 | +0.05 | +0.047 |
| New informationDoes it contain information absent from the stories covered in the last 7 days (list provided)? | 0.89 | 0.85 | 0.98 | 0.91 | 0.13 | +0.05 | +0.045 |
| 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. | 0.04 | 0.05 | 0.02 | 0.04 | 0.03 | −0.19 | −0.007 |
| OverclaimDo the headline or conclusions claim more than the body supports? | 0.14 | 0.30 | 0.03 | 0.16 | 0.27 | −0.11 | −0.017 |
| Sum of contributions | +0.694 | ||||||
| Claim-check penalty | 0.000 | ||||||
| Total, as published | 0.694 | ||||||
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/2026-11-18/22/"><img src="https://rubric.hrih.si/badge/2026-11-18/22.svg" alt="Ranked #22 on Rubric, 18 Nov 2026" width="245" height="28"></a>Claim check
Claim check: 8 no source given. Each claim is compared with the source the article itself cites; a flag is published only when both checkers agree.
No source givenNumber
Of the paper's 83 invariant clauses, 56 are tagged Tested, 12 Believed and 15 Open.
The article gives no link or citation for it.
The article says “Of the 83 invariant clauses, 56 are tagged Tested, 12 Believed and 15 Open”
The source says No source to compare with.
No source givenNumber
Applying the checklist to the author's harness, 11 of 30 items hold, 14 do not and 5 were not tested.
The article gives no link or citation for it.
The article says “of its 30 items, 11 hold, 14 do not and 5 were not tested (Section 7”
The source says No source to compare with.
No source givenNumber
Of the 30 cards, 24 rest on scripted providers alone in the recorded evidence.
The article gives no link or citation for it.
The article says “Of the 30 cards, 24 rest on scripted providers alone in the recorded evidence”
The source says No source to compare with.
No source givenNumber
In the author's one-shot mode, a command with a 3 s timeout was reported as timed out yet kept running, finished its side effect and held the turn for 45 s.
The article gives no link or citation for it.
The article says “in our harness’s one-shot mode a command with a 3 s timeout was reported as timed out, yet ran on, finished its side effect, and held the turn for 45…”
The source says No source to compare with.
No source givenNumber
In the author's harness, 41 of the 65 parity scenarios can run over the pipe, so the other 24 are checked in one transport only.
The article gives no link or citation for it.
The article says “in our harness 41 of the 65 parity scenarios can run over the pipe, so the other 24 are checked in one transport only”
The source says No source to compare with.
No source givenCausal conclusion
In ten real runs per option, a prompt sentence against repeating the closing did not lower the repeat rate, nor did rewording the tool's result; a display rule removed it.
The article gives no link or citation for it.
The article says “In ten real runs per option, a prompt sentence against repeating did not lower the rate, nor did also rewording the tool’s result; a display rule removed it”
The source says No source to compare with.
No source givenGeneralization
Across the ten other harnesses tested, none of those with a most permissive mode keeps the sandbox in it, and none checks at runtime a model's claim made with no call behind it.
The article gives no link or citation for it.
The article says “none of those with a most permissive mode keeps the sandbox in it; none checks at runtime a claim the model makes with no call behind it”
The source says No source to compare with.
No source givenNumber
A parity run of every scenario in two transports found 32 baselines and six test files of one transport failing unnoticed.
The article gives no link or citation for it.
The article says “a parity run of every scenario in two transports found 32 baselines and six test files of one transport failing unnoticed (case 3”
The source says No source to compare with.
Also covered by
No other coverage was clustered with this story.
Elsewhere
Elsewhere: HN 6 pts, not on front page
Recorded for context only, checked 18 Nov 2026, 05:47 UTC. It played no part in the ranking.