Pento: from agent harnesses to software factories — 2027 stays lights-on
Sept 24 from Leonardo Piñeyro (Pento): Stripe merged 1,300+ no-human-code PRs a week; Ramp’s agent writes 3 of 4 merged PRs. Factories sell the queue and gate, not the model. Autonomy only as wide as cheap verification; the lit factory keeps a person on merge while the gate moves upstream to specs and risk.
Leonardo Piñeyro’s Sept 24 essay at Pento is a layer map with a bet at the end. The opening receipts set the tempo: as of February 2026, Stripe was merging more than 1,300 pull requests a week that contained no human-written code; Ramp says its in-house agent now writes three of every four PRs it merges. Neither treats the agent as a chat window — work enters as a ticket or Slack message, and a pull request comes out for a person to review. That shape is the industry’s software factory: many agents on a queue, with a review gate between them and production. Factory and 8090 alone raised $285 million this year to sell one.
The middle of the piece sorts what teams keep conflating. A framework is building blocks you assemble (LangGraph, Agents SDK, Mastra). A harness is a finished agent you tailor — loop, tools, context, permissions, subagents (Claude Code, Codex, OpenCode, Pi). A platform hosts and governs either. LangChain’s coding agent moved from 52.8% to 66.5% on Terminal Bench 2.0 without changing the model underneath; the harness changed. When models sit this close, the wrapper often decides whether a long task finishes or drifts after fifty steps — hence the 2026 shorthand agent = model + harness.
Point that stack at a backlog instead of a user and you get the factory table that matters for review:
| Coding agent | Software factory | |
|---|---|---|
| Unit of work | A session | A backlog |
| Runs | One task at a time | Many in parallel |
| Engineer’s job | Steer, then read the diff | Spec, tune the loop, own the gate |
| Bottleneck | Attention | How fast a change can be verified |
Vendors split on the source of truth. Autonomous loops (Factory, Devin) aim ticket → PR with less human time in the loop. Spec-governed factories (8090, Pento’s Steer) keep the spec as the record and attach evidence — tests, screenshots, traces — to every pull request. What they sell is less the agent than the queue, sandboxes, specs, evals, and audit trail around it.
The live argument is the last step. In a lights-on factory, agents write most of the code and a person still reads the diff before merge. In a lights-off factory, code ships that no human has read. The public lights-off case is StrongDM’s “no human writes, no human reviews”; the public failure is Dex Horthy’s team abandoning a fully automated factory after months of quiet comprehension debt — green builds, codebase nobody still understands. Pento’s read: 2027 belongs to the lit factory. You can only hand a loop as much autonomy as you can cheaply verify; tests catch what they can express, not a design that gets a little worse with every green build. So the gate stays, and it moves upstream — from reading every diff to approving specs, architecture, and the changes that carry real risk (auth, payments, migrations). The Monday advice is narrow: get specs, tests, and evals in order; write down per repo which changes a person must read; open one queue where a passing suite is a fair definition of done.
Why it matters for code review
Always-on harnessed throughput beats peak-session heroics only when verification is the neck you widen, not the step you delete. A continuous factory that keeps the lights on where oracles are weak is the same claim as Osmani’s light-and-dark map and Kazzee’s order of operations — with the added industry scoreboard that the queue-and-gate product category is already funded. Compare ROBOCO on who actually merges and Winslow on coordinator-done is not merge.