A quarter of enterprise PRs are now AI-generated; review is the bottleneck
Greptile puts AI-generated PRs at 25% of enterprise volume, with revert rates roughly at human parity.
The real work is review capacity, business context, and orchestration, because model capability has become table stakes.
The review wall. AI-generated PRs are already about 25% of enterprise volume, with revert rates near human levels.
Greptile's median user opens 50 PRs a month, and the p90 hits 500.
Plan review checkpoints, automated validation, and agent-specific checklists before the volume outruns your team.
Context is the blocker. Business context is what blocks agent adoption.
An ex-Apple security engineer's ISA document ties every feature to evals and runs 1,500 automated checks hourly.
Conductor captures Slack, Discord, and meetings into a Postgres-backed agent so agents query company memory with SQL.
Document one workflow before buying more agent tooling.
Capability is table stakes. Model selection now runs on reliability and orchestration.
Opus 5.5 doubled reasoning tokens to 84,000 while output barely grew, and DeepMind's kernel demo used 93 subagents and 2 billion tokens.
Evaluate models by failure rate on your workflows, and treat orchestration as production infrastructure with cost caps and human checkpoints.
My take
I trust the volume signal more than the parity claim.
Greptile leaving fingerprints on a quarter of enterprise PRs means review has already changed.
But I'd hold the factory and orchestration visions at arm's length: they're vendor demos until they show production cost and failure data.
Instrument your own pipeline, document one workflow, and set human checkpoints before buying more agent tooling.
That's the org design decision you can act on this quarter.
The videos behind this brief
We watched these so you don't have to.
AI code has reached enterprise parity; review workflows are the bottleneckAI Engineer
Engineering becomes factory work: code less, tune agentsAI Engineer
AI builders are moving from factory automation to orchestrated, human-led agent teamsAI Engineer
The missing business context is the real agent adoption blockerDavid Ondrej
Model wave proves pacing is working: AI is getting less dumbTheo - t3․gg
Agent Teams Are the New Unit of Work; Governance Hasn't Caught UpAI EngineerVideo and thumbnails remain the property of their respective creators. tldw.news provides editorial analysis, commentary, and discovery links to original content.