#ai#team

AI-generated code is written quickly. Accepting it takes longer

GitHub, VS Code Copilot, and OpenClaw show that AI agent speed means little without review, security, and environment isolation.

  • One idea behind the different news stories
  • Diffs that cannot be split
  • Who holds the keys
  • A workspace for every agent

One idea behind the different news stories

GitHub redesigned the review interface in the Copilot app because some changes simply cannot be split into small pull requests: a major refactoring or migration has to arrive as one diff, and the discussion around that diff grows with it. During the same weeks, GitHub Enterprise gave teams a full export of all credentials, VS Code Copilot gained parallel chat sessions, and the weekly release introduced the experimental `/worktree` command for isolated agent workspaces.

Alongside this is the story of OpenClaw: one developer’s personal project grew to 388,000 stars, 81,000 forks, and 80,000 commits in ten months, and its maintainers are now rebuilding the process to avoid drowning in a stream of pull requests.

The different news stories share one idea: the faster agents write code, the more the surrounding pipeline costs companies—review, access control, and environment isolation.

Anyone who does not build this pipeline in advance ends up with an incident.

Diffs that cannot be split

Stacked pull requests are a practice of splitting a large change into a chain of small PRs so the reviewer can focus on one idea at a time. It works for new functionality. It does not work for a data schema migration or replacing a library throughout the repository: such a change either rolls out as a whole or not at all. GitHub redesigned the pull request viewer in the Copilot app for this case—diffs thousands of lines long and discussion threads with hundreds of comments should open as quickly as a small PR.

The agent is precisely what speeds up generating such a migration—but review interfaces have been designed for years around 200-line diffs, not 20,000-line ones.

Who holds the keys

GitHub Enterprise added an export of the organization’s complete credentials registry:

  • SSH keys
  • classic and fine-grained personal tokens
  • OAuth app tokens
  • GitHub App tokens—user-to-server
  • installation

One table for all enterprise owners and applications.

As soon as an agent is allowed to commit, open PRs, and call APIs, the number of active tokens in the organization grows faster than the number of people.

During an incident, the security team has minutes to determine which tokens have been compromised and what they can access.

Without an inventory, answering this question is guesswork.

KT.Team applies the same principle one step earlier: the LLM & Security Gateway limits what a model or agent can access before a token is issued.

Assess where AI can deliver impact in your process

clients@kt.team Telegram @kt_team_it

A workspace for every agent

  1. The July VS Code Copilot release added agent window management and parallel chat sessions, so you can conduct a review in one window while the agent works in another.

  2. The August weekly release went further: the experimental `/worktree` command creates an isolated working copy of the repository for the agent.

  3. The idea is simple: if one agent is handling a migration while another fixes a bug, they cannot share one file system because one agent’s changes will overwrite the other’s state.

  4. This is standard Git discipline for parallel work by multiple people, applied to parallel work by multiple agents—and without it, asynchronous development with several agents becomes a race to overwrite someone else’s file first.

Maintainers do not scale like GPUs

OpenClaw—a personal AI assistant that started as Peter’s weekend project

Steinberger in late 2025

As of August 26, 2026: 388,000 stars, 81,000 forks, and more than 80,000 commits. In interviews, the maintainers discuss not features but how to rebuild the review process for a flood of pull requests.

A company can quickly purchase more computing power to handle growing demand.

You cannot simply buy more attention from human maintainers who can thoughtfully read someone else’s code—it grows far more slowly than the workload. A project that does not prepare its review process in advance either slows growth or starts merging blindly.

Conclusion

TTU—time to use—is usually calculated for the tool itself: how quickly it delivers a result. These four news stories add the other half of the metric: how quickly a team can accept what the tool produces. An agent that delivers code in seconds is worthless if review takes a week, no one knows which tokens it has, its changes conflict with another agent’s changes, and the number of PRs grows faster than people can read them.

AI-native development is about engineering a pipeline that can withstand this speed: reviewing large diffs, inventorying access, isolating environments, and building a process for nonlinear growth. The simple outcome—a fast, safe merge—requires costly engineering work behind the scenes.

Discuss the article: AI writes code fast. Accepting it…

Enter your email or phone number so we can get back to you.

Send via:

By clicking the button, you accept the offer and grant consent to the processing of personal data.

Related Cases

All cases

Related solutions

All solutions
AI Accountant in Your Environment: Deployment and Inference CostAI Agent Platforms: Purpose and Open OptionsAI Agent Sprint for LeadersAI agent development: one process in 1–2 weeks

Articles on the topic

AI in Business Collection

Related videos

All videos
Shorts What “KPI” did Amazon set for developers in the AI era? AITeams & hiringManagement
Shorts Can't Keep Up with AI? Build Communication and Accountability AI
Shorts AI does not empower the weak! AI

News on the topic

All news