Radar, November 5, 2026: ten signals from X on Claude Code subagents, MCP, ACP and dev tools
In short: The standout signal of the day: Claude Code subagents use much less of the usage limit, according to @ClaudeDevs. Alongside it, ten short signals on Claude, Codex, MCP, the Agent Client Protocol and other dev tools, each with its source on X and, where possible, a link to the context in the briefing.
Illustration by Claude (AI), drawn with code: a radar screen with ten signals.
1.Claude Code subagents use much less of the limit
Confidence: high (own announcement; saving not independently measured)A running background subagent re-read the whole conversation every 30 seconds to update its status line. Haiku 5.5 now writes that line from short notes, which @ClaudeDevs says is about 99% cheaper per update. Anyone running many subagents can check whether the weekly limit lasts longer; it has not been measured independently yet.
Source on X: @ClaudeDevs
Context in the briefing: Subagents in the November 5 briefing
Read also: Claude Code news
2.Postman Desktop reportedly put plugins into Claude Code and Codex unasked
Confidence: low (single report)A user reports that Postman Desktop installed plugins in Claude Code and Codex without asking and automatically approved its own startup hook. It is a single report on X, not confirmed by Postman. To be safe, check which plugins and hooks are set up in your own environment.
Source on X: @foukation · @soysarcasme
Context in the briefing: Postman in the November 5 briefing
Read also: Skills and plugins · Codex news
3.Living Brain: a shared memory for ChatGPT, Claude and Cursor
Confidence: medium (announcement by the maker, not tested)ChatAndBuild launched Living Brain, a memory that AI apps such as ChatGPT, Claude and Cursor can read and edit, so context moves along to another tool. The post drew 123 likes and we have not tested the service. With any shared memory, pay attention to what data you put into it.
Source on X: @testingcatalog
Context in the briefing: Living Brain in the November 5 briefing
Read also: MCP and agent tooling
4.GitHub Copilot works as an agent in JetBrains Air through ACP
Confidence: lowAccording to a post, GitHub Copilot is available as an agent in JetBrains Air through the Agent Client Protocol (ACP), locally or in the cloud through Air Teams, on your Copilot subscription. The post has little reach and we have not checked the claim with JetBrains or GitHub.
Source on X: @8zero7
Context in the briefing: ACP in the November 5 briefing
Read also: MCP and agent tooling
5.Android Studio Rabbit 2 opens up to third-party agents
Confidence: lowAndroid Studio Rabbit 2 gets 'Bring Your Own Agent' and exposes project tools through the Agent Client Protocol. Another post mentions 40+ coding agents; that number is not confirmed. For Android developers it means the choice of agent is no longer tied to the editor.
Source on X: @AgentBreak40737 · @todddesil
Read also: MCP and agent tooling
6.Check of 61 official remote MCP servers: the auth spec is not followed everywhere
Confidence: lowAccording to @oauthpad, all 61 tested official remote MCP servers return a 401 and support PKCE S256, but only 25 offer Client ID Metadata Documents and 40 lack the iss parameter. Anyone building an MCP client should not expect the same login flow on every server. It is a single measurement by an account with little reach.
Source on X: @oauthpad
Context in the briefing: MCP auth in the November 5 briefing
Read also: MCP and agent tooling
7.MCPGuard looks for gaps in MCP SDKs
Confidence: lowResearchers published MCPGuard, a framework that automatically checks where MCP SDKs deviate from the specification and where validation checks are missing. The post drew 2 likes and we have not read the paper. The topic matters to anyone writing their own MCP server.
Source on X: @ayboxyz
Read also: MCP and agent tooling
8.Google Antigravity 2.29 gets a voice mode and remote agents
Confidence: mediumAccording to the changelog (via TestingCatalog), Antigravity v2.29.0 gets a voice mode based on Gemini Live, remote agents and browser updates. The post drew 226 likes; we have not read the changelog itself.
Source on X: @testingcatalog
Context in the briefing: Antigravity in the November 5 briefing
9.Computer use tool in the Gemini API is generally available
Confidence: high (own announcement)Google made the computer use tool in the Gemini API generally available, with no breaking changes from the preview, for browser, mobile and desktop agents. According to Google, anyone who already built on the preview does not need to change anything. The post drew 1,244 likes.
Source on X: @GoogleAIStudio
Context in the briefing: Computer use in the November 5 briefing
10.Cognition launches SWE-2 Flash and Factory acquires HumanLayer
Confidence: high (own announcements; benchmark not independently measured)Cognition introduced SWE-2 Flash, a fast coding model that the company says beats all US open-weight models on key evals, even ones three times its size. Factory is acquiring HumanLayer, a multiplayer IDE and cloud for coding agents. The benchmark claim comes from Cognition itself.
Source on X: @cognition · @factory
Context in the briefing: SWE-2 Flash in the November 5 briefing
