← All tech signals
Skills signal

The Agent Wrote the PR

Software engineering’s most tedious middle step, writing the code itself, is being handed to an AI agent that opens its own pull request against a real ticket queue.

What is it?

Routine software engineering — the tedious middle step of turning a ticket into working code — is being handed to autonomous agents that open their own pull requests against real backlogs. Revenue and commit share are compounding fast. The unpriced limit is review capacity, not code quality.

Why now?

Writing code was always bottlenecked by scarce engineering hours; models made fluent code generation cheap and fast, so that bottleneck loosens overnight — cognitive labour that once took a person days now runs in parallel, unattended. The real test is whether trust scales as fast as output does.

Who’s building it?

Model labs and coding-agent startups build the tools; enterprise engineering teams decide how much autonomy to grant them.

Claude CodeCognition / DevinGitHub CopilotEnterprise eng teamsReview tooling
The full read
Why now, who is building it, where the market goes, and what stays unsettled.
Get weekly →
01 · The market
Who’s in it
Builders — the coding-agent labs and platforms building the autonomy layer.
Buyers — engineering leaders paying for velocity, then discovering the review bill that follows.
Gatekeepers — platform owners and enterprises setting how much autonomy engineers may grant.
What’s at stake
Valuations already assume agents keep compounding output indefinitely — a bet that review capacity can somehow keep pace with it.
Engineering headcount shifts toward supervising code, not writing it.
A bad merge scales as fast as a good one.
What’s unsettled
Whether benchmark gains translate into fewer real production incidents.
How much autonomy a team grants before review is the bottleneck.
Whether value concentrates in the model layer, the agent product, or the platform.
02 · How the demand matureda progression, not a timeline
Novelty
Teams trial an agent on low-stakes tickets, curious more than committed.
Retention
Faster merges make the workflow sticky; going back feels like a demotion.
Willingness to pay
Business subscriptions multiply as agents are trusted with real backlog, not just demos.
Category
Engineering reorganises around supervision, and review capacity becomes the standing budget line.
03 · Where the market goesthe forks, not a prediction
→Supervision or substitution — whether engineers stay in the loop reviewing agent work, or agents review each other and humans step back further.
→Coding agents consolidate around one or two dominant platforms as enterprise trust concentrates.
→Review tooling becomes its own market, sized directly to the volume agents now produce.
→Commit share keeps climbing even where production quality hasn’t kept pace.
Everything here turns on one unmeasured thing: whether an agent-written pull request is actually as safe as a human-written one, at scale.
04 · The Agentic Delegation deskadjacent markets, moving now
Prompt Literacy
The skill market forming around directing agents.
AI-skill job listings jumped from 0.87% to 5.7%.
Forbes
The Agent Handoff
Agents now need a common protocol to work together.
Google’s A2A protocol launched with 50+ tech partners.
Google Developers Blog
The Four-Hour Firm
Delegated work lets one founder run a team.
1 in 3 new startups are now solo-founded.
Forbes
05 · The Wire
01
Anthropic says Claude Code run-rate revenue passed $2.5 billion
Anthropic (company announcement) ↗
02
Cognition says Devin’s annualised revenue run-rate reached $492 million
TechCrunch ↗
03
GitHub’s Copilot app runs parallel agent sessions in isolated worktrees
Help Net Security ↗
04
28.3% of agent-written pull requests merge instantly; many stall in review
arXiv (preprint, not peer reviewed) ↗
The panel on this signal
They agreed on
–Adoption and revenue are compounding faster than review practice is maturing to properly match them, at scale.
–Value is shifting from writing code to supervising it.
–Rules on how much autonomy to grant sit with enterprise buyers, not the agent makers.
The ML systems engineer disagreed on
–Commit volume and associated revenue figures measure throughput and cost, not reviewer confidence in code quality; faster merging under time pressure is consistent with looser review standards as much as with genuine trust in agent output.
Know more about the method →
✦How this signal was built
01 · Panel
A Delphi panel of 4 AI analysts, each reasoning independently.
→
02 · Debate
They challenge and revise each other’s reads over two rounds.
→
03 · Evidence
Every claim is tied to a cited, linked source on The Wire.
→
04 · Synthesis
A moderator writes it into one read, disagreement and all.