The same work on a third of the tokens: tuning a fleet of coding agents
Comments on “The same work on a third of the tokens: tuning a fleet of coding agents”
Claim-checked. Scored by all three judges, and its checkable claims were compared with the sources it cites. How far stories are checked
Alexander Marquardt reports that over ten weeks of Claude Code transcripts, his fleet of parallel coding agents cut tokens per changed line 3.3x (about 2.5x when weighted by token price) and merged 4.4 times as many PRs a week on 19% fewer tokens. Most tokens were cache reads of re-sent context, so the gains came from one-task agents with a context ceiling and fewer cache expiries. It is one project run by one person with changes landing together, and rework per PR rose slightly.
Why it ranked
No. 3 of 46 with a score of 0.71. The panel rated it highest on technical depth (0.92) and practical relevance (0.87), lowest on significance (0.56).
- AI judgeJudge 2claude-haiku-5-5
- Detailed, measured engineering writeup on token efficiency of coding-agent fleets with transcript-based methodology; single-author case study with confounded effects, so moderate significance and overclaim risk.
- AI judgeJudge 3gpt-5.6-luna
- A detailed, measured engineering account of reducing coding-agent token use across a real fleet, with transcript methodology, operational changes, caveats, quality metrics, and reproducible tooling. Its main limitation is single-project scope and partially confounded interventions.
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
0.71total under rubric v2.3
- Significance0.56
- Reader interest0.84
- Audience breadth0.57
- Practical relevance0.87
- Substantiated0.85
- Technical depth0.92
- New information0.86
- Counts against: Promotional0.04
- Counts against: Overclaim0.26
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 | 0.40 | 0.67 | 0.56 | 0.27 | +0.24 | +0.135 |
| Reader interestWould a well-informed senior engineer or tech founder want to read this today, and likely pass it on? | 0.81 | 0.82 | 0.90 | 0.84 | 0.09 | +0.26 | +0.219 |
| Audience breadthHow many people who build technology does it concern? | 0.54 | 0.50 | 0.67 | 0.57 | 0.17 | +0.13 | +0.074 |
| Practical relevanceHow useful is it to someone who builds software or tech products? | 0.93 | 0.80 | 0.88 | 0.87 | 0.13 | +0.22 | +0.191 |
| 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. | 0.99 | 0.70 | 0.86 | 0.85 | 0.29 | +0.05 | +0.043 |
| Technical depthHow technically substantive is it? | 1.00 | 0.85 | 0.90 | 0.92 | 0.15 | +0.05 | +0.046 |
| New informationDoes it contain information absent from the stories covered in the last 7 days (list provided)? | 0.81 | 0.90 | 0.88 | 0.86 | 0.09 | +0.05 | +0.043 |
| 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 | 0.04 | 0.04 | 0.01 | −0.19 | −0.008 |
| OverclaimDo the headline or conclusions claim more than the body supports? | 0.30 | 0.35 | 0.12 | 0.26 | 0.23 | −0.11 | −0.028 |
| Sum of contributions | +0.715 | ||||||
| Claim-check penalty | 0.000 | ||||||
| Total, as published | 0.715 | ||||||
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/2026-11-09/3/"><img src="https://rubric.hrih.si/badge/2026-11-09/3.svg" alt="Ranked #3 on Rubric, 9 Nov 2026" width="231" height="28"></a>Claim check
Claim check: 3 matches source · 5 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
Over ten weeks and six changes, the author's tokens per changed line of code fell 3.3x.
The article gives no link or citation for it.
The article says “Ten weeks and six changes later, tokens per changed line of code are down 3.3x”
The source says No source to compare with.
No source givenNumber
The agents now merge 4.4 times as many PRs a week on 19% fewer tokens.
The article gives no link or citation for it.
The article says “the agents now merge 4.4 times as many PRs a week, on 19% fewer tokens”
The source says No source to compare with.
Matches sourceQuote
Claude Code's documentation states that the model doesn't remember anything between requests, so Claude Code re-sends the full context.
The cited source says the same thing.
The article says “The model doesn’t remember anything between requests, so Claude Code re-sends the full context”
The source says “The model doesn’t remember anything between requests, so Claude Code re-sends the full context: the system prompt, your project context, every prior message and tool result, and your new message”
Source cited by the article: code.claude.com/docs/en/prompt-caching
No source givenNumber
Cache expiry while agents waited accounted for 31% of agent usage and fell to 0.7% after the wait cap and cache fixes.
The article gives no link or citation for it.
The article says “Cache expiry fell from 31% of agent usage to 0.7%”
The source says No source to compare with.
No source givenNumber
On the first day of the hook-enforced context ceiling, median context per request fell from 218k to 101k tokens.
The article gives no link or citation for it.
The article says “On the ceiling’s first day, median context per request, across all kinds of agent, fell from 218k to 101k tokens”
The source says No source to compare with.
Matches sourceNumber
Cache reads cost a tenth of the normal input price or less, and cache writes cost 1.25x for a five-minute cache or 2x for a one-hour cache.
The cited source says the same thing.
The article says “costs a tenth of the normal input price or less, depending on the model. Writing that part into the cache costs 1.25 times the normal price for a five-minute cache,…”
The source says “Cache writes: When new content is written to the cache (25 percent more than base input tokens for 5-minute TTL”
Source cited by the article: docs.claude.com/en/docs/build-with-claude/prompt-caching
No source givenNumber
The coordinator's tokens per merged PR fell to 7% of their baseline after a program took over routine fleet work.
The article gives no link or citation for it.
The article says “The coordinator’s tokens per merged PR fell to 7% of their baseline”
The source says No source to compare with.
Matches sourceQuote
Manus called the KV-cache hit rate the single most important metric for an agent.
The cited source says the same thing.
The article says “Manus called the KV-cache hit rate the single most important metric for an agent”
The source says “If I had to choose just one metric, I'd argue that the KV-cache hit rate is the single most important metric for a production-stage AI agent”
Source cited by the article: manus.im/blog/Context-Engineering-for-AI-Agents-Lessons-from-Building-Manus
Also covered by
No other coverage was clustered with this story.
Elsewhere
Elsewhere: HN 3 pts, not on front page
Recorded for context only, checked 9 Nov 2026, 05:46 UTC. It played no part in the ranking.