Projects coordinate fleets, models route themselves: AI coding tools shipping Sep. 4–11 (2026)

The qualifying releases in this issue fall between September 4 and September 11, 2026. The cutoff is 10:00 a.m. Pacific on September 11. Cursor introduced coordinator-driven Projects that manage thousands of subagents over long development cycles; GitHub Copilot shipped GPT-6 Astra, deprecated MAI-Code-1-Flash, added Project HydraFusion semantic routing, and launched centrally managed enterprise agent permissions; Claude Code synchronized gateway pricing to client cost telemetry while adding effort caps; JetBrains Junie brought Gemini 3.8 Flash with a 75% launch discount; and Devin automated child pull-request cleanup on session archiving.

Engineering managers and tooling leads can scan the summary table first, then use the review actions in each section to set up policy evaluations, model migrations, or infrastructure canaries.

The week in one view

Tool and release	Date	What changed	Why it matters	First review action
Cursor: Projects	Sep. 10	Cursor launched Projects in beta. A coordinator agent plans and delegates work to fleets of subagents, maintains persistent context across cloud and local machines, and runs scheduled or event-driven Subscriptions. 1	Long-horizon engineering moves from single prompt sessions to autonomous subagent swarms that retain learning over months.	Pilot one bounded migration in Projects, test shared-context file sync, and observe coordinator delegation behavior.
GitHub Copilot: model lifecycle and routing	Sep. 4–10	GPT-6 Astra reached general availability under usage-based billing, MAI-Code-1-Flash was deprecated across all Copilot surfaces on September 10 with MAI-Code-1.1-Flash as replacement, and Project HydraFusion entered `/experimental` semantic routing in Copilot CLI. 2 3 4	Teams gain a reasoning model built for self-verification, while legacy model paths retire and semantic routing automates cost-latency trade-offs.	Migrate any remaining MAI-Code-1 workflows to 1.1-Flash, evaluate GPT-6 Astra on long-horizon tasks, and test HydraFusion in CLI scripts.
GitHub Copilot: enterprise governance	Sep. 4–9	Enterprise-managed permissions arrived for agent operations in Copilot app, CLI, and VS Code Agent Host with un-overrideable deny/ask/allow rules. Copilot for JetBrains added managed sandbox policies in preview and `/ide` CLI connection, while VS Code 1.136 and 1.137 added Agent Merge and recurring scheduled tasks. 5 6 7	Central administrators can enforce strict shell, filesystem, and network boundaries across developer endpoints and automate recurring maintenance.	Configure an organization deny/ask policy in `.github-private/copilot/managed-settings.json` and verify that local auto-approvals cannot weaken it.
Claude Code 2.1.261–2.1.268	Sep. 4–10	Added Claude apps gateway pricing synchronization with client `/cost` and telemetry, `maxEffortLevel` caps across all providers, self-hosted runner session cleanup via `--remove-session-state`, a 1 GB disk cap on saved tool results, and fixes for symlink permissions and WebFetch timeouts. 8	Gateway administrators can align billing telemetry with actual enterprise rates, enforce global thinking-token ceilings, and prevent disk exhaustion on shared runners.	Deploy 2.1.268 to gateway-connected clients, verify `/cost` telemetry against gateway rates, and test `maxEffortLevel` in organizational policies.
JetBrains Junie: Gemini 3.8 Flash	Sep. 9	Gemini 3.8 Flash became available in the Junie IDE plugin and Junie CLI with a 75% launch discount. The model spends extra tokens on loop-based exploration, reproduction scripts, and automated test verification. 9	Teams gain an exploratory agent model tuned for multi-step tasks that beats larger frontier models on DeepSWE v1.1 while offering lower introductory token rates.	Benchmark 3.8 Flash against 3.7 Flash on a complex multi-file bug fix to measure whether the additional tokens reduce manual human review.
Devin: session lifecycle and team plugins	Sep. 9	Archiving a parent session now automatically offers to close child pull requests. Devin also added re-scanning on new commits with bulk remediation via MCP, installable team plugins with security notices, and Git Bash as the Windows default. 10	Multi-branch agent exploration leaves fewer orphaned pull requests, while packaged plugins simplify standardizing skills and MCP servers across teams.	Archive an exploratory session to test child PR closure, and package team-wide MCP tools into an enterprise-managed Devin plugin.

Cursor: Projects coordinate subagent fleets over extended timelines

Cursor introduced Projects in beta on September 10. Access is located in the left navigation panel. A Project coordinates long-running work that spans an entire feature, a multi-file framework migration, or a complete greenfield application. Rather than expecting a single conversational session to retain state, Projects introduce an architectural split between planning and implementation. 1

The coordinator agent in a Project plans the broader initiative, breaks the requirements into individual tasks, and delegates execution to subagents running in parallel. The coordinator does not edit source files directly. It schedules work, monitors completion, and returns the accumulated output for developer review. Because a Project executes on its own cloud machine, closing the local editor does not interrupt background progress. When a task requires execution or compilation on the developer's workstation, the coordinator spins up a local agent to handle that step on-device. 1

Projects address agent amnesia through a persistent shared context. Each Project maintains a synchronized directory of files across all cloud and local worker instances. As subagents explore the codebase, run tests, and discover build quirks, they record their findings into the project context. Subsequent subagents inherit those instructions, preventing repetitive onboarding turns. 1

The release also integrates Subscriptions. Developers can bind a coordinator agent to a Slack channel, an automated schedule, or incoming pull requests. For example, connecting a Project to a bug-triage channel allows the coordinator to wake automatically when an issue is reported, plan the investigation, and spawn diagnostic subagents without human prompting. 1

Tooling teams evaluating Projects should structure their first canary around three operational checkpoints:

Context growth and drift: Audit the size and precision of the shared files after several subagent turns to ensure obsolete assumptions are pruned rather than compounded.
Concurrency governance: Measure API consumption when coordinators trigger large subagent swarms against complex tasks.
Local worker boundaries: Confirm which operations execute on the cloud container versus the local workstation when subagents perform system-level tasks.

GitHub Copilot: model additions, deprecations, and adaptive routing

GitHub Copilot updated its model catalog and model routing infrastructure between September 4 and September 10. OpenAI's GPT-6 Astra became generally available on September 4 for Pro+, Max, Business, and Enterprise tiers across Visual Studio Code, Visual Studio, Copilot CLI, the coding agent, GitHub Mobile, JetBrains IDEs, Xcode, and Eclipse. The model is billed at provider list pricing under usage-based billing. 2

GPT-6 Astra emphasizes continuous verification. In GitHub's internal evaluation, the model plans iteratively, combines diagnostic commands with automated verification steps, and validates task criteria before returning control to the user. For enterprise administrators, Astra is governed through the Copilot model policy and is enabled automatically under standard default model enablement rules unless disabled globally. 2

Beside this addition, GitHub deprecated MAI-Code-1-Flash on September 10 across all Copilot experiences, including Chat, inline edits, agent workflows, and code completions. GitHub directs administrators and users to MAI-Code-1.1-Flash as the immediate replacement. Existing workflows calling MAI-Code-1-Flash must be repointed. Administrators can verify policy enablement in Copilot settings to make the 1.1 alternative visible in client selectors. 3

Copilot CLI added automated semantic model selection with Project HydraFusion, now available in /experimental. HydraFusion appears directly in the CLI model picker. When selected, it routes prompts dynamically across local, cloud, and compound models based on task complexity, balancing performance, execution latency, and token expenditure. 4

GitHub Copilot CLI model picker showing Project HydraFusion in preview — *The model selector in GitHub Copilot CLI lists Project HydraFusion in preview alongside foundation models, displaying context and reasoning capabilities. 4*

The September 7 release also introduced Jira integration to the Copilot app canvas, allowing developers to import Jira issues, select items for execution, and maintain ticket context across investigation and pull-request preparation. In Visual Studio Code 1.137, recurring agent tasks entered public preview, supporting hourly, daily, or weekly schedules. Experimental voice mode enables developers to speak to, interrupt, or steer running Copilot agents. 4

GitHub Copilot: enforceable enterprise governance across IDEs and agents

GitHub delivered central policy controls for agentic execution this week. On September 9, GitHub released enterprise-managed permissions for Copilot agent operations in general availability for Business and Enterprise tiers across the Copilot app, Copilot CLI, and Visual Studio Code sessions backed by Agent Host. 5

The feature establishes an administrative configuration file (.github-private/copilot/managed-settings.json) that enforces deny, ask, and allow rules across three operational dimensions:

Shell commands: Restrict specific executable binaries, flags, or script patterns.
Filesystem access: Define strict boundaries on file reads, writes, and modifications.
Network domains: Control outbound HTTP/HTTPS destinations that agents can contact during tool calls.

Managed policies cannot be bypassed or softened by developer workspace settings, local command-line flags, or auto-approval rules. Policies can also be segmented by team within the enterprise. 5

For JetBrains IDEs, GitHub released enterprise-managed sandboxes in public preview on September 8. When enabled by administrators, the policy centrally governs sandbox activation, filesystem boundaries, network egress, corporate proxy parameters, developer tool availability, and macOS Keychain access. Copilot locks the affected settings in the IDE and flags them as organizationally managed. 6

The same JetBrains update introduced cross-file cursor jumps for next edit suggestions, allowing developers to navigate multi-file edits in sequence, along with global project context in chat. Terminal Copilot CLI users can now use /ide to bridge their terminal session to an active JetBrains editor, sharing diagnostics, active file selections, and local Python virtual environments. 6

These releases follow the August 31 weekly release, which brought Agent Merge into public preview in Visual Studio Code 1.136 to resolve review feedback, failing CI checks, and merge conflicts automatically. 7

Claude Code 2.1.261–2.1.268: gateway alignment, effort caps, and runner hygiene

Anthropic shipped versions 2.1.261 through 2.1.268 of Claude Code between September 4 and September 10. The updates concentrate on enterprise gateway administration, resource limits, and session isolation. 8

Version 2.1.268 (September 10) added pricing: configuration support to gateway.yaml for Claude apps gateways. When defined, signed-in Claude Code clients automatically inherit those custom pricing rates through managed settings. This synchronizes client-side /cost calculations and spend telemetry directly with the organization's central billing meter. The release also added gateway startup warnings for unconfigured access_control.allow_cidrs, introduced the gatewayInternalNetworks managed setting to authorize logins from specific enterprise IPv4 blocks, and added --json support across all plugin lifecycle commands (install, uninstall, update, enable, disable). 8

For self-hosted runner infrastructure, 2.1.268 introduced claude self-hosted-runner --remove-session-state (off by default). This flag automatically deletes per-session directories under /_sessions/ when an agent finishes, preventing ephemeral build artifacts and temporary files from filling runner disks. The update also added a 300-second deadline to WebFetch (CLAUDE_CODE_WEBFETCH_DEADLINE_MS) to eliminate hangs on stalled HTTP streams, and fixed symlink permission bypasses on macOS and Linux where commands referencing real paths avoided rules written on symlink aliases. 8

Version 2.1.267 (September 9) introduced the maxEffortLevel setting, configurable globally or per model under modelSettings. This allows administrators to set an upper bound on extended thinking effort across all providers, including Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. The release also repaired prompt-cache stability during model switches, MCP server reconnects, and background worker forks. 8

Earlier in the cycle, version 2.1.265 (September 8) established a 1 GB cap on tool results saved to disk, supported pointing --plugin-dir at a folder containing multiple plugins, and added automatic archiving for Visual Studio Code sessions inactive for 14 days. Version 2.1.261 (September 4) raised maximum command and background task inline outputs (bashOutputMaxChars and taskOutputMaxChars) to 128,000 characters, added /skill-doctor to identify and remove unused skills, and surfaced organization policy errors on the /status line. 8

JetBrains Junie: Gemini 3.8 Flash balances token expenditure against verification

JetBrains published Gemini 3.8 Flash availability in Junie on September 9 for both the IDE plugin and Junie CLI. JetBrains is offering an introductory 75% discount on the model. 9

Gemini 3.8 Flash alters the token allocation strategy compared to Gemini 3.7 Flash. While 3.7 Flash focuses on fast, single-pass code generation, 3.8 Flash allocates more tokens toward iterative exploration: navigating directory structures, writing reproduction scripts for observed bugs, executing tests via tool calls, and verifying diffs prior to task completion. 9

Gemini 3.8 Flash DeepSWE v1.1 cost and resolve rate comparison chart — *The DeepSWE v1.1 evaluation compares Gemini 3.8 Flash against Claude Opus 5 and GPT-5.6 Luna on task resolution versus average cost. 9*

JetBrains references Google's DeepSWE v1.1 benchmark results, where Gemini 3.8 Flash achieves resolve rates competitive with larger frontier models while maintaining a lower average cost per resolved task. For routine, single-file edits or lightweight refactoring, JetBrains recommends remaining on Gemini 3.7 Flash to minimize per-token expenditure. 9

Devin: child session cleanup, bulk remediation, and packaged plugins

Cognition updated Devin on September 9 with workflow refinements for complex agent operations. When archiving a parent session that spawned sub-tasks, Devin now presents an archive prompt listing all open pull requests created by child sessions, allowing engineers to close them in bulk. This eliminates orphaned PR branches from abandoned explorations. 10

In sessions and across Devin MCP clients, developers can now re-run existing code scans against only new commits and trigger bulk remediation across multiple detected findings in a single turn. For team configuration, Devin introduced Plugins, which package skills, rules, hooks, and MCP server configurations into a single installable unit. Enterprise and organization administrators can mark plugins as required or optional, and installation surfaces a security notice itemizing the plugin's requested capabilities. 10

Additional updates in the September 9 release include defaulting to Git Bash on Windows environments, clickable diff paths in the Changes and PR tabs that open the embedded editor, touch controls for remote desktop on iPad, and real-time Slack and Microsoft Teams sync toggles accessible from the session sidebar. 10

Coverage notes

The Windsurf JetBrains plugin changelog route maintains version 2.12.27, dated August 17, 2026, as its latest published entry. An alternative check against the Windsurf blog was intercepted by a network security checkpoint, yielding no dated release for September 4–11. 11

Aider's official release history begins with an undated main branch section, while its GitHub releases page records v0.86.0 on August 9, 2026, as the latest dated release. Continue.dev's GitHub releases page and Atom release feed list v2.1.0-vscode from June 19, 2026, as the latest visible version. Neither tool published a date-verifiable release within this 7-day window. 12 13

Review before next Friday

Cursor: Run a trial project in the new Projects beta. Test coordinator task decomposition on a small multi-step refactor, verify shared-context file persistence between cloud and local agents, and audit Subscriptions triggers.
GitHub Copilot models: Repoint any scripts or workflows calling MAI-Code-1-Flash to MAI-Code-1.1-Flash following the September 10 deprecation. Test GPT-6 Astra under usage-based billing on complex tasks to measure autonomous verification quality.
GitHub Copilot governance: Draft a pilot .github-private/copilot/managed-settings.json policy to evaluate deny/ask rules for shell and network commands. In JetBrains IDEs, review the enterprise sandbox preview settings and test /ide CLI workspace bridging.
Claude Code: Update client installations to 2.1.268. If running a Claude apps gateway, configure pricing: in gateway.yaml to synchronize client /cost reporting. Test maxEffortLevel caps and enable --remove-session-state on self-hosted runners.
JetBrains Junie: Evaluate Gemini 3.8 Flash under the 75% introductory discount on multi-step bug investigations, comparing solve reliability against Gemini 3.7 Flash.
Devin: Verify child session PR closure when archiving exploratory tasks, and assess whether custom MCP servers and team rules should be consolidated into installable team plugins.
Windsurf, Aider, and Continue.dev: Continue monitoring their release repositories; no date-verifiable updates appeared during this window.