# Source record: Making delegation work (six-article series)

**Public copy prepared 2026-10-22.** Pins re-checked: the production study (arXiv 2512.04123v1) and the Intercom outcomes page were re-fetched and their figures and wording re-read on 2026-10-22; the six remaining URLs were re-fetched and their titles and publication dates confirmed the same day. Not re-checked: the body text of those six pages, which stands on the 2026-09-15 reading recorded below.

**Status:** source-and-claim record for an analytical series with illustrative examples. No Rethinking Agents benchmark, deployment or experiment was executed for any article. Every worked example, workflow, cost figure, patch and timing calculation is original and fictional unless a row below says otherwise.

**Supports:** [1 · How much autonomy does this job need?](/blog/how-much-autonomy-does-this-job-need) · [2 · A research assistant whose work you can use](/blog/a-research-assistant-whose-work-you-can-use) · [3 · The cheapest agent can be the most expensive workflow](/blog/the-cheapest-agent-can-be-the-most-expensive-workflow) · [4 · What does an agent need to remember?](/blog/what-does-an-agent-need-to-remember) · [5 · Your coding agent's reviewer needs evidence](/blog/your-coding-agents-reviewer-needs-evidence) · [6 · When more agents are worth it](/blog/when-more-agents-are-worth-it). Hub: [Delegation review kit](/resources/delegation-review-kit).

## Real sources

| ID | Source | Date | Checked | Supports | Limits |
|---|---|---|---|---|---|
| S1 | [Pan et al., *Measuring Agents in Production*, v1](https://arxiv.org/html/2512.04123v1) | 2025-12-02; fieldwork 2025-07-28 to 2025-10-29 | 2026-09-15; re-read 2026-10-22 | 306 survey responses, 20 case interviews; main analysis uses 86 deployed-or-pilot responses with question-specific denominators; authors describe limited autonomy and human evaluation | Self-report, convenience recruitment, selection toward successful deployments; no causal claim |
| S2 | [Intercom, *From resolutions to outcomes*](https://www.intercom.com/blog/from-resolutions-to-outcomes-evolving-how-fin-delivers-value/) | 2026-03-12 | 2026-09-15; re-read 2026-10-22 | An outcome counts when Fin completes the configured action, including a procedure that hands off to a human | Vendor pricing explanation; no ROI, resolution-rate or price claim carried forward |
| S3 | [Anthropic, *How we built our multi-agent research system*](https://www.anthropic.com/engineering/multi-agent-research-system) | 2025-06-13 | 2026-09-15; resolves 2026-10-22 | Parallel research architecture, internal evaluation uplift, substantial extra token use; decomposition and shared context as constraints | Internal evaluation, older models; uplift does not transfer to our designs |
| S4 | [METR, early-2025 experienced open-source developer study](https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/) | 2025-07-10 | 2026-09-15; resolves 2026-10-22 | 16 developers, 246 tasks; longer completion time with AI despite perceived improvement | Historical; always pair with S5; not a current estimate |
| S5 | [METR, developer-productivity design update](https://metr.org/blog/2026-02-24-uplift-update/) | 2026-02-24 | 2026-09-15; resolves 2026-10-22 | Selection effects, unreliable concurrent-work timing, weak evidence on later speedup magnitude | Supports uncertainty only |
| S6 | [Anthropic, *Effective harnesses for long-running agents*](https://www.anthropic.com/engineering/effective-harnesses-for-long-running-agents) | 2025-11-26 | 2026-09-15; resolves 2026-10-22 | Progress records, feature requirements and environment checks when resuming a coding task | Vendor coding case; does not validate our handoff design |
| S7 | [Anthropic, *Harness design for long-running apps*](https://www.anthropic.com/engineering/harness-design-long-running-apps) | 2026-03-24 | 2026-09-15; resolves 2026-10-22 | Generator/evaluator separation, structured handoffs; some context-reset machinery removed after a model change | Selected demonstrations; does not show agent reviewers replace human review |
| S8 | [Anthropic, *Demystifying evals for AI agents*](https://www.anthropic.com/engineering/demystifying-evals-for-ai-agents) | 2026-01-09 | 2026-09-15; resolves 2026-10-22 | Distinguishes tasks, repeated trials, graders and final outcomes | Stated method only; not independent validation |

## Per-article record

| Article | Sources | Evidence status | Required treatment | Fictional elements (keep labelled) |
|---|---|---|---|---|
| 1 · How much autonomy | S1, S2 | Analysis; worked example; nothing executed | Preserve S1's sample-selection and denominator caveats; no causal inference. S2 illustrates configured outcome semantics only | Docking-station supplier search, decision map, three stopping-point designs, approval packet, proposed evaluation checks |
| 2 · Research assistant | S3 | Analysis; complete fictional dossier; nothing executed | No transfer of S3's uplift to the fictional design; source ambiguity is distinguished from contradiction | Documentation-search decision, six invented source cards, claim register; the companion dossier is fiction in full |
| 3 · Cheapest agent | S4, S5 | Hypothetical inputs with reproducible arithmetic | No vendor-price or achieved-savings claim; preserve units and the accepted-task denominator; cite S4 only with S5 | Every workflow number: review and correction time, setup, maintenance, break-even volume, the $350-per-100-tasks example |
| 4 · What to remember | S6, S7 | Original handoff design; nothing executed | S6/S7 are specific engineering examples, not proof; the H01 branch is separate from the original dossier; no persistence or recovery trial claimed | The completed handoff record, continuation event H01, the interrupted action |
| 5 · Reviewer needs evidence | S7, S8 | Predicate analysis and expected cases; no execution | Preserve the UTC and millisecond assumptions; keep the logical finding distinct from test output; every expected-case row stays "Not run" | The date-filter patch, its defect, the review packet, the four boundary cases |
| 6 · More agents | S3 | Original decompositions; arithmetic checked | Elapsed and summed execution time differ (42 min serial, 21 elapsed parallel, 45 summed parallel under stated assumptions); neither establishes token cost or human effort | Both decompositions, the filled worker return, the schedule arithmetic, the explicitly hypothetical three-investigation comparison |

## Claim boundaries for the whole series

- Sources were read, not reproduced. "Reported by others" is the only evidence class present; "Observed by us" is empty for all six articles.
- The source set is disproportionately Anthropic-authored because those reports are unusually detailed; they do not represent the whole market.
- Fictional examples are design proposals. Do not quote their numbers as findings.
- Refresh before reuse: a later version of S1 before citing a specific figure; S2 before any claim about current billing; S4/S5 when redesigned results appear; S3, S6, S7, S8 before any quantitative comparison or implementation.