AI coding tools weekly: Claude adds self-hosted runners while Copilot adds model and spend controls (2026)

From July 31 at 10:00 through August 7 at 10:00 Pacific, coding agents gained more places to run and more controls around when they may act. Claude Code added self-hosted environments and cross-session messaging; GitHub Copilot added Kimi K3, tunable reasoning, and comment-triggered cloud work; Cursor connected agents to Gmail, Drive, and Calendar. The practical question is shifting from "which assistant writes better code?" to "where can it run, what can it reach, and what will the run cost?"

The week in one view

Tool	Release date	What changed	Why engineering teams should care
Claude Code 2.1.221–2.1.224	Aug. 4–7	Self-hosted environments, cross-session messaging, a removed 200-agent session cap, stricter sandbox fixes, worktree isolation, and changes to background sessions, plugins, and review commands. 1	The executor boundary and several unattended-work defaults changed together. Upgrade tests need to cover permissions, concurrency, network access, and where work is committed.
GitHub Copilot with Kimi K3	Aug. 6	GitHub began gradually rolling out the open-weight model across Copilot clients. It is hosted by GitHub on Fireworks AI, billed at provider list pricing, and off by default for Business and Enterprise. 2	Model choice now includes an open-weight governance decision, an administrator policy, and usage-based spend.
Copilot cloud agent	Aug. 3	Users can select a reasoning level alongside a model. Higher reasoning can improve complex tasks but consumes more tokens and credits. 3	Teams can tune cost and depth per task instead of treating the model as the only quality control.
Copilot cloud agent automations	Aug. 3	A configured issue or pull-request comment can trigger documentation work, error investigation, or follow-up task creation. 4	The issue tracker and review thread become agent entry points. That makes trigger naming, approval, and duplicate-run handling part of platform design.
Cursor Google Workspace plugins	Aug. 3	New plugins let agents search and change Google Drive files, search and send Gmail, and read and update Calendar events. 5	The capability is useful, but the announcement does not specify detailed permission scopes. Treat it as an integration review, not a harmless plugin install.
GitHub Code Quality	Aug. 7	Enabling Code Quality no longer creates a ruleset that automatically requests Copilot review; GitHub turned off the settings it had created in matching existing rulesets. 6	A default that once created review traffic has been withdrawn. Teams relying on automatic review must check their rulesets rather than assume the behavior remains.
Copilot model roster	Jul. 31	GitHub announced September 1 deprecations for several Gemini, Claude, and Raptor Mini models, with a limited annual-plan exception for individual Sonnet 4.6 subscribers. 7	Enterprise model policies and pinned integrations need a migration rehearsal before the deadline.
GitHub Spark	Aug. 4	Spark stopped accepting new users and new apps. Existing users can access it through Aug. 31 to export apps; deployed apps continue to work, but `llm()` calls already stopped when GitHub Models retired on Jul. 30. 8	Prototypes have an export deadline and, if they use `llm()`, an inference-provider migration.
Devin / Windsurf-derived desktop	Aug. 1 and 5	Devin’s release notes added easier recovery for failed environment snapshots and a Linear reconnect action. Devin Desktop v3.6.27 restored Windows certificate-store loading and stopped local edit tools from writing through symlinks. 9 10	The changes are small but operationally important for corporate proxies, environment recovery, and preventing an approved edit from reaching an unexpected file.

Two cost-and-governance changes sit beside the feature releases. GitHub retired the Copilot Billing Preview app on August 4 and moved spend review to billing settings, which now expose AI-usage grouping, user-level budgets, cost centers, usage-pool allocation, exports, and the billing API. 11 Enterprise administrators also gained team-specific managed settings, with non-overridable enterprise values still winning while some plugins and marketplaces can be added at team level. 12

The agent is becoming a deployable service

Claude Code 2.1.224 is the clearest change in this direction. Team and Enterprise users can run Claude Code web, mobile, and desktop sessions on their own machines or containers through claude self-hosted-runner. The same release adds cross-session messaging on macOS and Linux, removes the 200-subagent-per-session cap, and fixes several sandbox and permission edge cases. 1

That is more than another front-end option. A self-hosted runner moves the question of where the agent executes from Anthropic’s service boundary into the team’s own network, images, credentials, patching, and observability. The release does not remove the need for those controls; it makes them the customer’s operating surface.

The releases immediately before it change the behavior of unattended work. Version 2.1.221 makes background sessions commit and push when appropriate, follow the repository’s CLAUDE.md instructions, and report where the work lives. Forked sessions now create their own worktrees. Version 2.1.222 applies isolation to destructive Git commands and prevents repository-local settings from enabling Remote Control; that switch must be enabled at user scope. 1

GitHub is approaching the same service boundary from the other side. A Copilot cloud agent automation can start from a precise comment on an issue or pull request. A team might use a comment to investigate a stack trace, generate documentation, or create follow-up work. The feature is available on paid plans that include the cloud agent, but Business and Enterprise organizations still need an administrator to enable the cloud-agent policy. 4

Cursor’s Google Workspace plugins push the boundary beyond the repository. The announcement says agents can search, open, create, and organize Drive content; search and send Gmail; and read, create, or update Calendar events. It does not spell out narrower permission scopes or a detailed data-handling boundary. 5

The common pattern is easy to see: an agent can now be started from a web or mobile surface, a code comment, a team-owned machine, or an external work system. The rollout unit is no longer just an IDE extension. It is a service with identity, network reach, storage, audit logs, and a failure owner.

Model choice now exposes the bill

GitHub’s Kimi K3 rollout adds another model to nearly every Copilot surface, including the cloud agent, CLI, app, GitHub.com, mobile, JetBrains, Xcode, and Eclipse. GitHub hosts the open-weight model on Fireworks AI and says the rollout is gradual. Business and Enterprise administrators must enable its policy; if they do not, the model remains unavailable to the organization. 2

The release also puts Kimi K3 under provider list pricing rather than hiding it inside a flat seat comparison. That makes a model pilot inseparable from a budget policy: who can select it, for which tasks, with what usage cap, and against which accepted-patch baseline.

Copilot cloud agent now exposes a second control. Users can choose a reasoning level alongside the model, and GitHub warns that higher reasoning consumes more tokens and credits. 3 This is a useful control because implementation, debugging, and review do not need the same depth. It is also a new source of variance in cost and latency: a test that records only the model name is now incomplete.

The September 1 model deprecations make the migration deadline concrete. GitHub lists Gemini 3.1 Pro, Claude Opus 4.5 and 4.6, Claude Sonnet 4.5 and 4.6, and Raptor Mini for removal across Copilot Chat, inline edits, ask and agent modes, and code completions. The suggested replacements vary by model. Claude Sonnet 4.6 remains available to individual subscribers on annual plans, but that exception does not cover the general enterprise migration path. 7

The right pilot ledger now has at least six fields: selected model, reasoning level, task type, successful patch or review outcome, latency, and total credits or provider spend. Vendor benchmark scores can sit beside that ledger, but they should not replace it.

Defaults are moving in both directions

GitHub reversed one of its own automation defaults this week. When Code Quality became generally available on July 20, enabling it created a ruleset that automatically requested Copilot review on pull requests, including new pushes and draft pull requests. On August 7, GitHub stopped creating that behavior and turned off the settings in rulesets it had created, unless a customer had edited the ruleset. Teams can still configure automatic Copilot review themselves. 6

This is a small product correction with a large lesson for rollout plans: defaults are policy. If a workflow depends on them, pin the desired behavior in a ruleset or configuration file and test what happens after the vendor changes its setup path.

The same Code Quality surface gained a more constructive automation. In public preview, it can start an agent that creates a pull request containing a repository-specific code-coverage workflow. The generated workflow builds the code, runs tests, produces a coverage report, and uploads it with least-privilege permissions. 13 The important boundary is preserved: the agent proposes a pull request; the repository still has a review and merge step.

GitHub’s team-specialized managed settings point in the opposite direction from a global default. Enterprise administrators can now target teams with itemized configuration files. Enterprise-only values remain authoritative, while marked-overridable values and additive plugin or marketplace settings can vary by team. 12 This supports staged rollouts, but the least-restrictive merge behavior for users in multiple teams deserves a policy test of its own.

The sharpest boundary change is GitHub Spark’s retirement. New users and new apps stopped on August 4. Existing users have until August 31 to export app code, while deployed apps continue to run. Apps that call llm() need a separate inference provider because GitHub Models no longer serves that function. 8 This is not a routine feature deprecation; it changes both the editing surface and, for some apps, the runtime dependency.

What to review before next Friday

Claude Code: Stage 2.1.221–2.1.224 against wrappers that use background sessions, nested agents, Remote Control, plugins, MCP, and worktrees. Record the selected model, network prompts, concurrency, and the final branch or pull request.
Self-hosted execution: For Claude’s runner or any comparable internal runner, define the machine image, credential injection, outbound network policy, patch cadence, session logs, and owner for a failed or abandoned session before opening a pilot.
Copilot models: Test Kimi K3, the current default, and the planned replacement for each model scheduled for September deprecation on the same task set. Include reasoning level, credits, latency, accepted patches, and review rework.
Comment-triggered agents: Reserve a small set of explicit trigger phrases, decide whether comments can start duplicate runs, and make the resulting agent status visible to the pull-request owner.
External plugins: Review Cursor’s Google Workspace access with the same care as an MCP integration. Start with a test account and document what the agent can read, create, send, or change; the announcement does not answer those questions for you.
GitHub defaults: Inspect Code Quality rulesets, Copilot cloud-agent policy, team-specific managed settings, plugin marketplaces, and user budgets. Do not assume last week’s default still exists.
Spark: Export any app that still matters before August 31. Search for llm() and replace it before treating the deployment as migrated.
Devin Desktop: If corporate TLS inspection or shared workspaces are in scope, test certificate loading and symlink handling on the supported desktop builds.

Aider’s official history and Continue.dev’s official release list showed no release dated inside this window. 14 15 No standalone Windsurf release with a readable, date-verifiable first-party entry was added as a forced row. The edition keeps the absence visible instead of turning an older release into a current update.

The week’s changes point in one direction, but not toward a single winning assistant. Agents are spreading across runners, comments, external workspaces, and mobile clients while model depth, permissions, and spend become explicit controls. Engineering teams can compare tools more honestly when their test plan measures the whole operating path rather than the code completion in isolation.