PO People & operating model
The PM and engineer role, team shape, economics, ROI, adoption.
From the issues
A Merge Is a Signature, Not a Test Result
No. GitHub's Octoverse 2026 reports 90% of agent-started pull requests merged within two weeks, a figure GitHub reports about its own platform. Merging measures acceptance, not correctness: in Foresight-SWE, models avoided the bug someone later fixed only 14.84% of the time. Track escaped defects per merged agent PR, not merge rate.
More Pull Requests, Same Finished Work: 718 firms, one flat line
Do coding agents raise real output? Activity, yes. Finished work, not measurably. A Harvard study of 718 firms found agents lifted pull requests 23% while resolved Jira issues showed no significant change and review time rose 3.45 days. Count finished work, not PRs, before calling it a productivity gain.
Stop Budgeting Tokens Per Developer: five questions that tell you what your AI spend bought
Budget AI by cost per delivered outcome, not tokens per developer. Gartner found only 13% of 2,002 CIOs report significant value from AI, and only 31% set kill criteria for pilots. Answer five questions each quarter, starting with what one escaped bug costs you.
The Review Bill: 21% more work, 91% more checking
Yes, agents raise output, and the review bill rises faster. Bain's survey of 293 senior technology leaders found developers complete roughly 21% more tasks while review time rises about 91%. Plan reviewer hours the way you plan sprint capacity, before you buy another agent seat.
The Last Human Gate: when automating governance adds work
Before an LLM runs a governance review gate, check whether a rules script already does the job. Across 899 runs on 300 synthetic projects, the best model passed 94.98% of gates. A deterministic baseline passed all 1,700. The human hours hide in that last 5%.
1,000 PRs in a week: five questions the headline didn't answer
A PR count on its own proves nothing about AI productivity. Kiro reports 1,000 merged PRs in 7 days with 3 engineers and 50+ agent sessions, self-reported, with no rollback rate, review time, PR size or escaped defects. Ask for the denominator before you clap for the number.
Numbers on this track
| Number | What it measures | Source | Checked | Track · issue |
|---|---|---|---|---|
| 94.98% | Best strict governance-gate success by an LLM (Gemini 3.8 Flash; GPT-5.6 Luna 83.29%, DeepSeek v4.1 Flash 74.18%; 300 synthetic projects, 899 runs) | Canale, arXiv:2609.29345 | ✓verified at source | POPeople & operating model Issue #01 |
| 1,700 of 1,700 | Governance gates passed by a deterministic rules baseline | Canale, arXiv:2609.29345 | ✓verified at source | POPeople & operating model Issue #01 |
| ~21% / ~91% | More tasks completed / more review time, per Bain's survey of senior technology leaders | Bain Technology Report 2026 | self-reported | POPeople & operating model Issue #02 |
| 47% | More concurrent workstreams, per the same Bain survey | Bain Technology Report 2026 | self-reported | POPeople & operating model Issue #02 |
| 718 | Firms in the Harvard study of coding-agent adoption | Chen and Stratton (Harvard), PDF | ✓verified at source | POPeople & operating model Issue #03 |
| +23% | Pull requests after adopting agents on top of assistants (resolved Jira issues: not statistically significant). Data from vendor Jellyfish's platform | Chen and Stratton (Harvard), PDF | ✓verified at source | POPeople & operating model Issue #03 |
| +3.45 days (+49%) | Increase in time from PR to merge after adopting agents | Chen and Stratton (Harvard), PDF | ✓verified at source | POPeople & operating model Issue #03 |
| 13% | Share of 2,002 CIOs reporting significant or transformational value from AI | Gartner CIO survey | ✓verified at source | POPeople & operating model Issue #03 |
| 31% | Share of organisations that have kill criteria for AI pilots (those that do are 2.3x more likely to be among the winners) | Gartner CIO survey | ✓verified at source | POPeople & operating model Issue #03 |
| 432 / 61% | Scenarios simulated, and the share of agent budget spent on finding bugs in the optimal teams | Herbold, arXiv:2610.20394 | ✓verified at source | POPeople & operating model Issue #03 |
| $150 to $1,200 | Per-developer monthly AI spend at two of 27 workplaces interviewed (anonymised) | Pragmatic Engineer (Orosz) | ✓verified at source | POPeople & operating model Issue #03 |
| 83% to 90% | Share of agent-started PRs merged within two weeks, February to September (above about 86% for others) | GitHub Octoverse 2026 | self-reported | POPeople & operating model Issue #04 |
| more than 40% | Share of September 2026 PRs that were agent-started, up from 17% in March ("agent-started" is a detection floor) | GitHub Octoverse 2026 | self-reported | POPeople & operating model Issue #04 |
| 81% | Share of teams that merged AI-written code or tests they did not fully understand (sample size not stated) | TestMu AI survey | self-reported | POPeople & operating model Issue #04 |
Papers on this track
- The Last Human Gate: Forward Deployed Engineering for Governance Automation arXiv:2609.29345
- What Should a Software Team Spend on Agents? Simulation-Based Cost Optimization of Mixed AI-Human Teams arXiv:2610.20394
Other tracks
ISIntent & specs CKContext & knowledge ABAgents that build UMUnderstanding & modernisation VTVerification & trust
Last updated . Reuse with credit under CC BY 4.0: “The Dabbawala Protocol, thedabbawalaprotocol.com”.