Market Pulse — Issue 38

Review-Shaped Work

The Infrastructure for Arguing Back

By Sevda Polat— July 15, 2026

Feature image for article: The Infrastructure for Arguing Back

Microsoft's coding-agent study reports a 24% increase in merged pull requests across tens of thousands of engineers, then the authors quietly concede that merged PRs are an output proxy, not delivered value. That concession matters more than the number. It points toward a structural question the usual conversation about agent adoption keeps skating past: the predictor of durable adoption may not be task complexity or model capability, but whether a domain already has organizational machinery for catching mistakes, contesting outputs, and reversing course before the cost lands on whoever is least equipped to absorb it.

Review-Shaped Work

The Infrastructure for Arguing Back

By Sevda Polat— July 15, 2026

Microsoft's coding-agent study reports a 24% increase in merged pull requests across tens of thousands of engineers, then the authors quietly concede that merged PRs are an output proxy, not delivered value. That concession matters more than the number. It points toward a structural question the usual conversation about agent adoption keeps skating past: the predictor of durable adoption may not be task complexity or model capability, but whether a domain already has organizational machinery for catching mistakes, contesting outputs, and reversing course before the cost lands on whoever is least equipped to absorb it.

Review Readiness Map

Agents produce output. Whether a domain can actually absorb that output, challenge it, trace it back, fix it when it's wrong, depends on review infrastructure that mostly predates the agent era.

Some workflows have spent decades building exactly this machinery. Diffs get reviewed. Transactions get disputed. Cases get escalated. Others have almost nothing standing between a generated number and the slide deck it lands in.

That gap matters more than most deployment conversations acknowledge, because standing up review mechanisms with real institutional weight takes far longer than deploying the agents that need them. Six domains, mapped by how ready their existing plumbing is to catch agent mistakes.

Review Readiness Map

Agents produce output. Whether a domain can actually absorb that output, challenge it, trace it back, fix it when it's wrong, depends on review infrastructure that mostly predates the agent era.

Some workflows have spent decades building exactly this machinery. Diffs get reviewed. Transactions get disputed. Cases get escalated. Others have almost nothing standing between a generated number and the slide deck it lands in.

That gap matters more than most deployment conversations acknowledge, because standing up review mechanisms with real institutional weight takes far longer than deploying the agents that need them. Six domains, mapped by how ready their existing plumbing is to catch agent mistakes.

Code Review

Code Has the Review Muscle Already Built

Review-readiness: High. Diffs, CI pipelines, tests, ownership rules, rollback. The full stack predates agents entirely. But Microsoft's deployment study flags a telling gap: adopters merged ~24% more PRs, yet merged PRs measure throughput, not value. The review surface works. What "good" means still doesn't.

Agentic Payments

Payment Rails Were Built for Exactly This

Review-readiness: High. Authorization, tokenized credentials, spending limits, chargebacks, Visa's 120-day confirmation window. Decades of dispute machinery. The hole: when user consent and merchant processing are both valid but the agent misread what the user actually wanted. No dispute code covers that yet.

Customer Support

Escalation Paths Exist, Case Memory Does Not

Review-readiness: Medium. Queues, escalation tiers, QA sampling all exist. Context handoff is where things fall apart. Nubank reports a 37-point NPS gain; a Taobao experiment finds shorter chats with worse ratings. Containment numbers can look great while the record handed to the next human gets thinner.

Enterprise Analytics

Agent Numbers Reach Decisions With No Checkpoint

Review-readiness: Low. No query lineage, no metric definitions, no caveats attached. Numbers land in decks and dashboards and become organizational fact. EntSQL's best system hit 15.9% accuracy on real enterprise questions. Plausible-sounding answers failing the actual business question, with nothing between generation and belief.

Internal Tools

Some Connectors Are Governed, Many Are Bare

Review-readiness: Varies. Enterprise workflows with approval chains sit alongside tool surfaces exposed via MCP that carry no authentication, no mandate, no audit trail. A 2026 survey found 40.55% of remote MCP servers exposed tools without authentication. The protocol's authorization layer remains optional by spec.

Web Access

Sites Recognize Agents but Lack Shared Terms

Review-readiness: Low-Medium. Cloudflare's search/agent/training taxonomy gives sites new classification levers. But classification alone isn't review. A fingerprinting study showed agents bypassing anti-bot systems while remaining distinguishable from humans. Sites can tell what's visiting. They just can't agree on what to do about it.

Coding Agent Evidence

Adoption and Impact of Command-Line AI Coding Agents: A Study of Microsoft's Early 2026 Rollout of Claude Code and GitHub Copilot CLI

Across tens of thousands of Microsoft engineers, coding-agent adopters merged roughly 24% more PRs. Retention tracked coding activity, not demographics. The authors are careful to note merged PRs measure output volume, not delivered value.

Why PRs specifically?

Agents gain traction where artifacts already flow through diffs, tests, CI, and rollback, making machine-generated work institutionally legible.

Does more output mean more value?

The authors say no, and that distinction matters: review systems make agent work visible, but visibility and quality are different things.

The Shift to Agentic AI: Evidence from Codex

Active Codex users grew more than fivefold in H1 2026. Skills adoption hit 26.6% in three months, and over 10% of users ran three or more concurrent agents weekly, pointing toward structured, repeatable delegation patterns.

What do skills signal?

Users encoding repeatable workflows look a lot like people building review-shaped contracts for agent output, not just prompting ad hoc.

How broad is multi-agent use really?

The heaviest concurrent-agent figures skew toward OpenAI's own teams; external enterprise adoption remains thinner and less even.

Coding Agent Evidence

Adoption and Impact of Command-Line AI Coding Agents: A Study of Microsoft's Early 2026 Rollout of Claude Code and GitHub Copilot CLI

Across tens of thousands of Microsoft engineers, coding-agent adopters merged roughly 24% more PRs. Retention tracked coding activity, not demographics. The authors are careful to note merged PRs measure output volume, not delivered value.

Why PRs specifically?

Agents gain traction where artifacts already flow through diffs, tests, CI, and rollback, making machine-generated work institutionally legible.

Does more output mean more value?

The authors say no, and that distinction matters: review systems make agent work visible, but visibility and quality are different things.

Coding Agent Evidence

The Shift to Agentic AI: Evidence from Codex

Active Codex users grew more than fivefold in H1 2026. Skills adoption hit 26.6% in three months, and over 10% of users ran three or more concurrent agents weekly, pointing toward structured, repeatable delegation patterns.

What do skills signal?

Users encoding repeatable workflows look a lot like people building review-shaped contracts for agent output, not just prompting ad hoc.

How broad is multi-agent use really?

The heaviest concurrent-agent figures skew toward OpenAI's own teams; external enterprise adoption remains thinner and less even.

The Unreviewed Number

The Number Arrives Looking Like an Answer

The EntSQL benchmark asked text-to-SQL systems to answer enterprise questions requiring domain knowledge: metric definitions, reporting conventions, organizational rules. The best evaluated English-input system reached 15.9% accuracy. Set that aside for a moment and consider what the other 84% looks like once it ships.

In coding, wrong output hits a diff, a test suite, a reviewer. Distrust already has a workflow. In analytics, the query produces rows. Rows become a number, and the number becomes a sentence in a deck. Somewhere between the SQL and the sentence, the metric definition, the filters, the caveats all fall away. Nobody later argues with the query. They argue with the conclusion, long after anyone could reconstruct whether "active customers" included trials, paused accounts, or subsidiaries.

A bad database mutation tends to be loud. A bad number can be quiet and durable. It looks precise enough to travel.

The Unreviewed Number

The Number Arrives Looking Like an Answer

The EntSQL benchmark asked text-to-SQL systems to answer enterprise questions requiring domain knowledge: metric definitions, reporting conventions, organizational rules. The best evaluated English-input system reached 15.9% accuracy. Set that aside for a moment and consider what the other 84% looks like once it ships.

In coding, wrong output hits a diff, a test suite, a reviewer. Distrust already has a workflow. In analytics, the query produces rows. Rows become a number, and the number becomes a sentence in a deck. Somewhere between the SQL and the sentence, the metric definition, the filters, the caveats all fall away. Nobody later argues with the query. They argue with the conclusion, long after anyone could reconstruct whether "active customers" included trials, paused accounts, or subsidiaries.

A bad database mutation tends to be loud. A bad number can be quiet and durable. It looks precise enough to travel.

TAKE NOTE

The coding contrast: Diffs, tests, CI, code review, ownership, rollback — all native to the domain. Analytics has none of these by default.

Why 15.9% isn't an outlier: On Spider 2.0, an o1-preview agent solved 17.0% of enterprise tasks vs. 91.2% on the toy benchmark. The drop is consistent.

Where meaning rots: Terms like "churn," "net retention," or "open incidents" can return rows before the organization has resolved what the term includes.

What would need to travel with the number: Metric definition and version, source tables and filters, executed SQL, known exclusions, reviewer decision, downstream use.

Where agents move fastest: Domains where the work product is already shaped for challenge. Analytics isn't one of them.

Primary Sources

Murphy-Hill, Butler, and Savelieva, 'Adoption and Impact of Command-Line AI Coding Agents,' arXiv, July 2026

Large-scale rollout showing 24% more merged PRs among adopters, though merged PRs measure output volume, not delivered value.

Johnston et al., 'The Shift to Agentic AI: Evidence from Codex,' arXiv, June 2026

Fivefold active-user growth and multi-agent concurrency suggest coding's review infrastructure is what lets agent work scale.

Gupta et al., 'Building Customer Support AI Agents at 100M-User Scale,' KDD 2026

Wang et al., 'Agentic AI and Human-in-the-Loop Interventions,' arXiv, May 2026

Liao et al., 'EntSQL,' arXiv, revised July 2026

Apostolou, Bosch, and Holmström Olsson, 'Agentic AI in Industry,' arXiv, May 2026

Primary Sources

Murphy-Hill, Butler, and Savelieva, 'Adoption and Impact of Command-Line AI Coding Agents,' arXiv, July 2026Large-scale rollout showing 24% more merged PRs among adopters, though merged PRs measure output volume, not delivered value.

Johnston et al., 'The Shift to Agentic AI: Evidence from Codex,' arXiv, June 2026Fivefold active-user growth and multi-agent concurrency suggest coding's review infrastructure is what lets agent work scale.

Quick links

Gupta et al., 'Building Customer Support AI Agents at 100M-User Scale,' KDD 2026

Wang et al., 'Agentic AI and Human-in-the-Loop Interventions,' arXiv, May 2026

Liao et al., 'EntSQL,' arXiv, revised July 2026

Apostolou, Bosch, and Holmström Olsson, 'Agentic AI in Industry,' arXiv, May 2026

Past Articles

The Middle Case

Visa's new agentic commerce rules hold cardholders responsible for agent-initiated payments "as if the Cardholder initia...

The Way They Move

Researchers at UC Davis watched seven commercial browsing agents navigate an instrumented website this spring. A behavio...

The Furniture Is Arriving Before the House Is Built

Across the agent ecosystem, multiple players are simultaneously shipping the same category of development: admin console...

What Governance Governs

Companies with governance tooling deploy twelve times more AI projects to production. Only 4 of 13 frontier-autonomy age...

Past Articles