Aethora Labs Inc.
All briefings
Editorial briefingSeptember 2, 2026 5 min read

Agents write faster than teams can review

Throughput is genuinely doubling in real deployments. The constraint did not disappear, it moved into review, and most teams are not instrumented for where it went.

AI agentsdeveloper productivitycode reviewMCP

For two years the argument about AI coding tools has been about whether the speedup is real. That argument is largely settled, and it settled in an awkward place: the speedup is real, it is large, and it lands almost entirely on the part of the pipeline that was never the bottleneck. What follows is what the recent evidence actually says, and what it implies if you are running an engineering team right now.

The clearest evidence comes from a company that mandated it

A July 2026 preprint, AI Writes Faster Than Humans Can Review, is the most useful data point published so far, because it studies a company that committed to a specific target rather than one that merely bought licences. The authors tracked 802 developers and 196,212 pull requests from January 2024 to April 2026 at a mid-sized firm that had committed to doubling merged pull requests per engineer since mid-2025.

It worked. Per-capita throughput reached 2.09 times the pre-mandate baseline by April 2026, which the authors describe as among the largest gains reported from a field deployment of AI coding tools. Merge and revert rates held steady, so this was not a case of shipping obvious garbage.

Two details matter more than the headline. First, per-reviewer load roughly doubled, and automated review overtook human review. The mandate did not just change how code was written, it restructured who or what approves it. Second, the authors are careful about causality: adoption and usage intensity were not randomly assigned, so they read the result as implicating an adoption-and-use channel rather than claiming exact causal attribution. That caution is worth copying when you present numbers like these internally.

The queue does not vanish, it relocates

Faros AI's research on the productivity paradox, drawing on telemetry from over 10,000 developers across 1,255 teams, points the same direction from a different angle. Developers on high-adoption teams completed 21 percent more tasks and merged 98 percent more pull requests. Over the same period, pull request review time rose 91 percent, average pull request size grew 154 percent, and bugs per developer rose 9 percent.

The finding that should give leaders pause is the last one: across overall throughput, DORA metrics, and quality indicators, Faros observed no significant correlation between AI adoption and improvement at the company level. Team-level gains did not aggregate. Larger changes sitting in longer queues is a coherent explanation, and it is the shape you would predict if the new capacity is being absorbed downstream rather than converted into delivered work.

Orchestration is becoming infrastructure

The tooling has already moved on from single-assistant workflows. In August 2026 AWS open sourced Kiro Crew, an orchestration layer for asynchronous coding agents aimed at long-running engineering workflows rather than single-turn edits. Analysts quoted in the coverage flagged governance and interoperability as the open problems, which is the same conclusion the throughput data arrives at from the other end.

JetBrains published its own research on agent adoption in the same month, alongside a unified control plane for managing agent-driven development across tools. When vendors start shipping control planes, it is a reliable signal that customers have more agents than they can currently account for.

The plumbing is quietly standardising

Underneath the orchestration story, agent access to real systems is becoming a first-class, supported surface. Microsoft made the Azure DevOps Remote MCP Server generally available on 21 August 2026, offering a hosted endpoint into work items, repositories, and pipelines with nothing to install. Days earlier, MongoDB announced a managed MCP server for Atlas that gives coding agents direct access to live application data.

This is the part most teams are underestimating. Standardised, hosted access to production systems is what turns an agent from a code generator into something that acts on your business. It is also what makes the review question a security question rather than a style question.

What this means for your team

If your AI programme reports acceptance rates or lines generated, you are measuring the part that is no longer scarce. Three adjustments follow from the evidence above.

Instrument review, not authorship. Review latency, pull request size, and queue depth are where the cost is now showing up, and all three are measurable today. If pull request size is climbing, gains are being converted into review debt rather than delivered software.

Treat automated review as a real pipeline stage. In the 2x study, automated review overtook human review as a matter of fact, not policy. If that is happening to you, it deserves the same scrutiny, ownership, and failure budget as any other gate, rather than arriving by accident.

Be honest about aggregation. Team-level wins that do not show up in company-level delivery are not fake, but they are not yet value either. That gap is usually a process problem sitting downstream of engineering, and no amount of additional agent capacity will close it.

If you want help turning this into something operational, our work on data monetisation starts from the same premise: the constraint is rarely the model, it is the pipeline around it. The live signal dashboard tracks the developments above as they land.

Sources

Put this into practice

Aethora Labs helps teams turn these signals into shipped automation.

AI automation solutionsAgentic workflows, RAG knowledge systems, and custom model integration for enterprise teams.Engagement & pricingSee how Aethora scopes and prices automation programs for teams of every size.Book a strategy callTalk through your use case with our team and map a path to production automation.The Aethora Signal archiveBrowse every briefing on AI and software-development developments.

Written and edited by the Aethora Labs team

This briefing was researched and written by a person, not by the autonomous editor. Every figure and claim is attributed to a named source in the list above.