Copilot Lets a Model Approve Safe Commands, Claude Code Stops Asking for Credits on Long Sessions, and Cursor Ships an Agent SDK

At a glance

• Copilot assisted permissions are generally available in Copilot CLI, the Copilot app and VS Code, and admins can block Allow All while keeping Assisted.
• Copilot CLI now offers Assisted without experimental mode, and GPT-6.1 Terra is rolling out in Copilot.
• Claude Code now compacts and keeps going when a standard context session runs past 200K tokens, instead of stopping with a 1M context credits error.
• Cursor's Agent SDK lets you define an agent as a folder of code, test it with evals and deploy it to Cursor or your own infrastructure.

The useful changes today decide what your agent may run without asking and how far it can go before it stalls. Copilot's news landed between 3:15 PM and 10:16 PM ET Monday, Claude Code shipped at 3:47 PM ET, and the Cursor Agent SDK entry appeared on the changelog after our Monday morning check, dated 20 October.

Treat today as a trust settings day: decide whether Assisted belongs in your Copilot policy, update Claude Code if long sessions keep stopping, and try the Cursor Agent SDK on one small bot before you plan around it.

Top Stories

1. Copilot assisted permissions go GA, and Copilot CLI v1.0.104 reaches latest

GitHub announced at 6:31 PM ET Monday that assisted permissions are generally available in Copilot CLI, the Copilot app and VS Code. When Copilot asks to run a command, reach a file system path or open a URL, a language model evaluates the request, so requests it classifies as clearly safe proceed on their own while everything else still waits for your approval. In the CLI you turn it on with /permissions, and the Copilot app and VS Code each add it to their permissions picker.

Enterprise admins get real controls with it. Managed settings can block Allow All while keeping Assisted available, disable Assisted separately, and set the default mode for new sessions, and if that default is unavailable or blocked by policy the session starts with manual approval. Copilot CLI v1.0.104 followed to npm latest at 10:16 PM ET, making Assisted Permissions available without experimental mode. Earlier, at 3:15 PM ET, GitHub made GPT-6.1 Terra generally available in Copilot on paid plans, pitched for everyday agentic coding at a lower cost than GPT-6.1 Sol, with a gradual rollout.

Practical dev impact: Assisted is a middle ground between approving every command and Allow All, so admins should pick a default mode now rather than letting each developer decide. Because Terra is enabled automatically unless your org turned off the global default, check your model policy if you track usage based spend.

2. Claude Code v2.1.311 keeps long sessions running instead of asking for credits

Claude Code v2.1.311 reached npm latest and next at 3:47 PM ET Monday, while stable stays on v2.1.308. The fix most people will feel is for standard context sessions that stopped with "Usage credits required for 1M context" once a request ran over 200K tokens. Claude Code now compacts the conversation and continues. A related fix stops Claude from seeing stale MEMORY.md and CLAUDE.md content right after automatic compaction when those files were edited earlier in the session.

Performance and agent fixes ride along. Long freezes at startup and after file edits are gone when a CLAUDE.md file holds a very long list item, Write and Edit no longer stall for up to 5 seconds on changes of thousands of lines, and interrupting Claude now takes effect even while an MCP server has not answered a resource read. Subagents can now use MCP tools from servers that finish connecting after the subagent starts, and SubagentStop hooks run once when a subagent finishes instead of about twice a minute.

Practical dev impact: If long refactors or log reviews in Claude Code kept stopping near 200K tokens, update to v2.1.311 and expect compaction instead of a credits prompt. Teams with SubagentStop hooks should check that their hook counts drop to one per finished subagent.

3. Cursor's Agent SDK turns a folder of code into a deployable agent

Cursor's main changelog now leads with the Agent SDK, dated 20 October but not on the page at our Monday 5:49 AM ET check. It is generally available and lets you define an agent as a folder of code, test it with evals, and deploy it to Cursor or your own infrastructure. In that folder, instructions.md is the prompt, each file in tools/ is one typed action, skills/ holds procedures loaded on demand, mcp-connections/ adds MCP tools, and channels/ defines where the bot receives messages.

Agents wake on pull request, comment and CI events from GitHub, GitLab and Bitbucket, on Slack mentions and DMs, on webhooks and on a cron schedule, and follow ups on the same pull request or Slack thread resume the same session. Cursor's docs install the CLI with npm install --global @grok/agent-sdk, then agent-sdk init, agent-sdk dev, agent-sdk eval and agent-sdk deploy. Templates include a security reviewer and a PR autofixer. The public npm registry still returns 404 for @cursor/agent-sdk, so do not run the changelog's npx @cursor/agent-sdk line. Use the docs install instead.

Practical dev impact: If you have been hand rolling a PR review or triage bot, this gives you evals and replayable GitHub events in one package, so start with one template on a test repo. Our read: treat the 20 October changelog date as real and the late appearance on the page as a visibility lag, not a fresh ship today.

Practical Impact Analysis

All three stories hand more autonomy to agents while adding guardrails, which is why they belong together today. Copilot lets a model approve safe commands but gives admins the final say on defaults, Claude Code lets sessions run longer by compacting instead of stopping, and Cursor packages agents with evals so you can see what a bot did before it runs on real pull requests. Copilot leads because the permission default you pick affects every developer in your org, while the Agent SDK is a build decision you can make this week.

Elsewhere at our check, Codex latest stayed on v0.164.0 while its alpha tag moved to v0.166.0-alpha.1 with no notes, and Gemini CLI latest v0.65.0 and preview v0.66.0-preview.0 were unchanged apart from a new nightly. The Cursor CLI changelog still leads with v2026.10.22.

Tutorial

Read the script before you run it.

1. Run step 1 on any machine with Copilot CLI or Claude Code.
2. Run step 2 inside a Claude Code project folder.
3. Run step 3 only if you want to try the Cursor Agent SDK in a scratch folder, using the install Cursor's docs show.

# We have not run this script end to end on every platform. It only prints and checks unless you uncomment a line.
# 1. Which versions are out, and which do you have? At 6:02 AM ET Tue 27 Oct npm showed
#    @github/copilot latest v1.0.104 and @anthropic-ai/claude-code latest v2.1.311, stable v2.1.308.
npm view @github/copilot dist-tags --json
npm view @anthropic-ai/claude-code dist-tags --json
command -v copilot >/dev/null && copilot --version || echo "no copilot CLI on PATH"
command -v claude >/dev/null && claude --version || echo "no claude CLI on PATH"
#    Optional npm installs only. Pinned at 6:02 AM ET Tue 27 Oct, latest channel:
# npm install -g @github/copilot@1.0.104
# npm install -g @anthropic-ai/claude-code@2.1.311

# 2. Very long list items in CLAUDE.md caused the freezes v2.1.311 fixes. List lines over 2000 characters.
for f in CLAUDE.md .claude/CLAUDE.md; do [ -f "$f" ] && awk -v f="$f" 'length > 2000 {print f": line "NR" has "length" chars"}' "$f"; done
echo "CLAUDE.md check done"

# 3. Cursor Agent SDK per docs (not the changelog npx line). At 6:02 AM ET Tue 27 Oct
#    @grok/agent-sdk latest was v0.3.0. @cursor/agent-sdk still 404s on the public registry.
npm view @grok/agent-sdk version
npm view @cursor/agent-sdk version 2>/dev/null || echo "not on the public npm registry; follow cursor.com/docs/agent-sdk"
#    Optional, npm install only. Pinned at 6:02 AM ET Tue 27 Oct, latest channel:
# npm install -g @grok/agent-sdk@0.3.0
#    Then from a scratch folder: agent-sdk init ./my-bot && agent-sdk dev
echo "Done. Nothing was installed or changed."

To try assisted permissions in Copilot CLI, start a session, run /permissions, and select assisted. Then ask it to list files and to delete a scratch folder, and watch which request waits for you.

Sources:

• GitHub Copilot assisted permissions generally available
• GPT-6.1 Terra in GitHub Copilot
• Copilot CLI v1.0.104 release
• Claude Code changelog
• Claude Code on npm
• Cursor Agent SDK changelog entry
• Cursor Agent SDK docs
• Codex releases
• Gemini CLI v0.67.0 nightly 20261027
• Cursor changelog
• Cursor CLI changelog

Recommended AI prompt

Copy this paragraph into ChatGPT, Claude, Gemini, Grok, or whatever AI you use.