Method
How the analyses behind the series were done: Chapter I's scoring of every kind of work, and Chapter III's look at how development organisations build and hire.
Chapter I: how every kind of work was scored
This lists every kind of work and every score. Click a row to see why each score is what it is: the case for, the case against and the sources, one score at a time. You can also sort any column and filter by area.
Chapter III: how the Forge analyses were done
These are my own analyses, done with Claude in September 2026. For each, this says how the organisations or items were chosen, what was counted, and what could make the number wrong.
Frontier experience at 40 implementers
Which 40. I started from the grantees of five funders that back effective development work: GiveWell (grants from 2022), Coefficient Giving (global health and wellbeing grants from 2022), AIM (its incubated charities), the Gates Foundation (global health and global development grants from 2022) and Prevail (all 11 known grantees). Each funder's list was shuffled with a fixed seed (20260917) and walked in that order, keeping the first 8 organisations that passed three screens:
- it hadn't already been drawn for an earlier funder;
- it's an implementer: it delivers a programme, service or product directly to people or frontline workers in low- or middle-income countries, rather than being a university, evaluator, multilateral, government, funder, consultancy, or technology provider building for others;
- 25 to 3,000 people list it as their employer on LinkedIn (at least 10 for AIM, whose charities are young).
| Funder | Grantees on the list | Screened | Passed | Not an implementer | Wrong size | Already drawn |
|---|---|---|---|---|---|---|
| GiveWell | 115 | 31 | 8 | 20 | 3 | 0 |
| Coefficient Giving | 79 | 24 | 8 | 14 | 1 | 1 |
| AIM | 54 | 47 | 8 | 27 | 10 | 2 |
| Gates Foundation | 1,342 | 135 | 8 | 121 | 5 | 1 |
| Prevail | 11 | 11 | 8 (of 10 passing) | 1 | 0 | 0 |
The draw can be rerun from the files, and it reproduces.
What counts. A top tech or AI company is one of about 200: the 18 frontier labs (every developer with a model in the LMArena top 10 since January 2025, plus the frontier developers in Epoch AI's data), every company on the Forbes AI 50 in 2024, 2025 or 2026, AI-native unicorns founded since 2015 (from Dealroom and CB Insights), and big tech (Meta, Google, Amazon, Microsoft, Apple and NVIDIA). A technical job is engineering, research, data science, machine learning or product, judged by the title held there. Only current staff count, not board members, advisers, volunteers or interns, and not roles under three months. "Since ChatGPT" means the role continued after 30 November 2022.
How it was counted. A LinkedIn people search (through LinkdAPI) for anyone listing one of the 40 as their current employer and one of the companies as a past employer, then the full profile of every hit, classified by Claude against the rules above. There were 18 hits. 3 had held a technical job at one of the companies, 1 of them since ChatGPT, and none had worked at a frontier lab.
What could make it wrong. "About 16,000" is the number of people who list the 40 on LinkedIn (16,368), not the organisations' own headcount. LinkedIn misses people who don't list their employer or experience. 17 of the companies had no usable LinkedIn page and weren't searched.
Frontier experience among those who help
Who. The four organisations most active in helping development organisations with technology: Agency Fund, IDinsight, Project Tech4Dev and Renaissance Philanthropy (counted across its funds).
How it was counted. The same search and rules as for the 40 found 5 current staff at the four organisations who had held a technical job at one of the companies, 3 of them since ChatGPT; 2 of the 5 aren't in technical roles now.
What could make it wrong. "About 700" is the number of people who list the four as their current employer on LinkedIn (708 on 28 September 2026: IDinsight 455, Renaissance Philanthropy 157, Project Tech4Dev 64, Agency Fund 32), measured as for the 40. Adding Renaissance Philanthropy's fund pages and Tech4Dev's Glific page takes it to 744. LinkedIn misses people who don't list their employer or experience.
The speed-up
The companies. AI-native product companies that have published a measured before-and-after figure for output per engineer. Intercom: merged pull requests per R&D employee (all of R&D, not only engineers) tripled in the 16 months to April 2026, its CTO wrote. PostHog: pull requests went from 1,441 in January 2026 to 4,725 in June while engineering headcount grew 10%, on its blog; 3x per engineer is my division. Replit: 2.9x the code from the same authors, January to June 2026, on its blog. Linear: 47% more merged pull requests per author from January to March 2026, as reported by LeadDev. Larger companies report similar gains: Block 2.5x code changes per engineer from January to mid-April 2026, Coinbase 78% more pull requests per engineer year on year, Robinhood 50% more commits per engineer since the start of 2025.
The development organisations. The same 18 organisations as below. Merged pull requests per active engineer in their main product repositories, bots and AI-agent accounts left out, grew 1.2x from December 2024 to April 2026 (three-month averages). Over the companies' shorter windows, January to June 2026, they grew about as little: 1.1x in pull requests, and 1.3x in lines of code for the same authors.
What could make it wrong. The companies' figures are their own and unaudited, and each measures output a little differently, over windows of 2 to 16 months. They're the companies that chose to publish, and they may have steered towards the number they published. The development organisations' figure covers the longest window, which flatters them. Counted in lines of code rather than pull requests, their engineers' output roughly doubled from 2024 to 2026, because pull requests got bigger; but lines of code grew faster at the companies too.
Engineers with AI agents in their code
Which organisations. Eighteen digital-first development organisations with public code: Dimagi, IDinsight, OpenFn, Ona, Glific, Simprints, Medic, ODK, KoboToolbox, DHIS2, OpenMRS, Avni, OpenCRVS, mSupply, Dalgo, Praekelt, the Agency Fund and Tech4Dev's AI projects. Ten AI-native product companies, chosen by hand: PostHog, Twenty, n8n, Zed, ToolJet, Lightdash, Activepieces, Plane, Label Studio and Cal.com. Each is venture-backed, builds its main product in public, has at least five regular engineers, and builds with or around AI.
What was counted. Engineers, not organisations. An engineer counts if they committed in at least four of the six months of the first half of 2024 and were still at the same organisation in the first half of 2026, with at least 20 commits. For each, I added up the lines of code and tests they changed across every public repository of their organisation, leaving out bots, docs, data, lockfiles and generated files, and capping any one commit at 5,000 lines. An engineer counts as having AI agents in their code if at least one of their 2026 commits is co-signed by one; almost all these signatures are Claude Code's. Growth compares the group's total lines in the first half of 2026 with the first half of 2024. Dimagi's Open Chat Studio figure is main-branch commits in 2026 to 15 September that carry a Claude co-author trailer.
What could make it wrong. Many tools, including Copilot and Cursor, don't sign commits, so some engineers who use AI show no sign of it. Keen engineers may take up agents and grow anyway, so this shows who ships more, not why. Lines of code aren't features, and AI-written code is wordier. Dimagi has 22 of the 65 engineers with agents in their code; without it they grew 1.9x against 1.3x, and 52% show no sign of them. The result holds with a threshold of 1 or 50 commits instead of 20, with AI use judged over 12 months, with engineers followed from 2025, and with any one organisation left out. It only sees public code, but engineers' private contribution counts show that work moving to private repositories doesn't explain the gap.
Job ads
Which ads. For development nonprofits, every 2026 engineering ad I could find, from careers pages, job boards and copies on aggregator sites (about half the texts), for a named list of organisations plus others found by search. These aren't the random 40. The chart uses 55 ads at 29 organisations, including Dimagi and leaving out OpenFn, Tech4Dev and Code for America; without Dimagi it's 5 of 52 (10%). For tech companies, 200 ads drawn at random (seed 20260917) from all 2,845 engineering ads live on 52 companies' Greenhouse and Ashby boards on 15 September 2026.
The rule. Does the ad ask for, expect or describe the hire using AI coding tools in their own engineering work? AI as the product, company boilerplate about AI, or a note about the hiring process doesn't count. Claude read in full every ad that mentioned AI at all, plus a random third of the rest (none of which met the rule).
What could make it wrong. The base is small: 8 of 55 has a 95% interval of roughly 7% to 27%. Ads are a weak proxy for practice. The nonprofit ads ran across 2026, while the companies' were live on one day.
Three ways
This is a judgement, not a count. I set out the three ways, and the forms outside help could take, from research on who supplies technical capacity to development organisations, what has been tried and stopped, and who might fund it, and weighed each on distribution, money, the incentive to be good, handing ownership to the implementer, and speed to start.