A Quiet Update With a Large Surface Area
On June 10–11, 2026, Anthropic pushed an update to Claude Code that most engineering teams probably skimmed past: nested sub-agents. Claude Code can now spawn its own sub-agents, up to five levels deep, allowing a single agentic session to coordinate multiple parallel workstreams across a codebase simultaneously.
It sounds like an internal architecture detail. It isn't. It's a signal about the scale at which AI is now operating inside real codebases — and it surfaces a review problem that most teams haven't adequately prepared for.
The Numbers Behind the Shift
To understand why this matters, start with a statistic from Anthropic's own 2026 Agentic Coding Trends Report: 78% of Claude Code sessions now involve multi-file edits. A year ago, that figure was 34%. The jump is not incremental — it represents a fundamental change in how AI coding agents operate.
AI is no longer suggesting a function in a single file. It's making coordinated changes across services, updating shared libraries, adjusting configurations, and touching interfaces that span multiple repositories in a single working session. The nested sub-agent capability accelerates this further: one top-level agent can now delegate to specialized sub-agents, each handling a discrete part of the codebase, and synthesize their output into a unified result.
For teams building microservice architectures or large monorepos split across many repositories, this means the blast radius of a single AI-assisted session is wider than it has ever been.
The Cross-Repo Review Problem This Creates
When AI writes code across services in one session, the resulting Pull Requests don't stay in one repository. A feature that touches an API gateway, a data service, and a frontend SDK will produce PRs in three different repositories — potentially across GitHub and GitLab if your organization uses both providers.
The conventional code review workflow was not designed for this. Opening each repository individually, finding the relevant PR, loading its diff, and mentally stitching together the cross-service context is cognitive work that scales poorly as both the number of repositories and the frequency of AI-generated changes increase.
This isn't just about inconvenience. When reviewers evaluate a PR in isolation, they're missing the full picture. A change that looks safe in one service may introduce a subtle contract violation in another. Without cross-repo visibility, that risk is invisible until it surfaces in production.
The Anthropic report captures a related tension: developers use AI in roughly 60% of their work, but can fully delegate only 0–20% of tasks. The gap exists precisely because high-stakes, cross-cutting changes still require deep human judgment. That judgment is harder to apply when the review interface fragments the picture across dozens of browser tabs.
What Engineering Leaders Should Watch
The nested sub-agent capability is a preview of where agentic coding is heading: more autonomous, longer-running sessions that make broader changes across larger codebases. Anthropic's report describes sessions stretching from minutes to hours, with agents pausing only for strategic human checkpoints.
For engineering leaders, the relevant question isn't whether to adopt these tools — adoption is already happening. The question is whether your review process can keep up with the surface area these tools create.
Three things worth examining now:
Cross-repo PR aggregation. If reviewers can't see all open PRs across repositories in one view, they're making review decisions with incomplete context. This becomes a compounding problem as agentic sessions span more services.
Risk signals at the PR level. When a single AI-assisted session produces multiple PRs across multiple repos simultaneously, prioritizing which ones need the most human attention requires heuristic signals — diff size, sensitive file modifications, merge conflicts — evaluated consistently across all of them, not one at a time.
Reviewer context, not just reviewer bandwidth. The bottleneck isn't always time. Often it's context. Reviewers need to understand how a PR fits into what else is moving across the organization at the same moment. That context is invisible inside a single-repo PR view.
The teams that navigate the agentic coding era well won't be the ones with the fastest AI generation. They'll be the ones with the clearest human oversight layer sitting above it.
Closing Take
Nested sub-agents in Claude Code represent a meaningful step toward AI systems that can autonomously coordinate complex, multi-service changes. Combined with the data point that 78% of sessions already span multiple files, it's clear that the review surface for engineering teams is expanding faster than most review tooling has adapted.
The answer isn't to slow down AI adoption. It's to raise the visibility layer that sits above it. When AI-generated changes span repositories, the engineering teams with the clearest cross-repo picture will make the best review decisions. If your team is starting to feel the pressure of distributed PRs across GitHub and GitLab, Code Board is built precisely for this moment — aggregating every open PR across every repo into a single, AI-powered board so the full picture is always in view.