--- title: 'In the News: November 12, 2026 (Evening)' description: "Claude Code 2.1.323 assumes a 1M context window for model IDs it does not recognize, which can break sessions on other vendors' smaller models." canonical_url: 'https://darkfactory.dev/news/2026-11-12-evening' markdown_url: 'https://darkfactory.dev/news/2026-11-12-evening.md' collection: news date_published: '2026-11-12T19:12:00-05:00' date_modified: '2026-11-12T19:12:00-05:00' --- # In the News: November 12, 2026 (Evening) Claude Code 2.1.323 now assumes a 1M context window for any model ID it does not recognize, and Anthropic's notes say sessions on other vendors' smaller models can fail every turn once they fill. The same release adds an experimental Code Mode and makes several unattended-run settings stop or ask where they used to proceed. Anthropic also opened a private beta of Claude Managed Agents that runs entirely in the customer's environment. ## 1. Claude Code 2.1.323 assumes a 1M context window for unrecognized models and adds an experimental Code Mode **[Claude Code v2.1.323](https://github.com/anthropics/claude-code/releases/tag/v2.1.323)** · Anthropic · GitHub release notes, November 12, 2026 The release notes are long. Four groups of changes bear on anyone running Claude Code without a person at the keyboard. The context window default changed. Claude Code used to assume 200K for a model ID it did not recognize and now assumes 1M. The notes say to use /autocompact to compact earlier, or to set CLAUDE_CODE_DISABLE_1M_CONTEXT=1 to keep 200K. They add a warning for sessions on other vendors' models with a smaller window: "once full, they can fail every turn until /clear". Code Mode is new. The notes describe it in one line, as "an experimental opt-in in which Claude writes JavaScript to call its tools", and point to a documentation page for turning it on, or turning it off for an organization. That page returned "Page Not Found" when we checked on the evening of November 12, so the one line is all we have read about how it works. Several settings now stop or ask where they used to proceed. If auto mode is requested and unavailable, a session used to start in a mode that could allow more; "It now starts in Manual mode, which asks first". If the settings that turn the sandbox on fail validation, claude -p and the Agent SDK used to run Bash unsandboxed; "Claude Code now stops at start-up and names the invalid setting". When a PreToolUse hook approves a tool call and the permission check then fails with an error, Claude Code decides the call as if no hook had answered. The tool that finishes an MCP server sign-in asks for confirmation even in bypass and auto modes. Sessions can also say much more to each other. The number of messages one session can send to your other sessions on the same machine without a prompt from you went "from about 20 to 100,000". Separately, the v1 MCP client is removed; MCP_SDK_GENERATION=v1 now runs the v2 client in a compatibility mode. We read the full notes. We did not install the release or test any of these behaviors, and the notes give no figures on how often the fixed cases occurred. **Why it matters:** If you point Claude Code at another vendor's model through a gateway, the window it assumes just grew fivefold, and Anthropic's own notes say such a session can fail every turn once it fills. Set CLAUDE_CODE_DISABLE_1M_CONTEXT=1 on those runners before the update reaches them. A headless job with an invalid sandbox setting now stops at start-up where it used to run unsandboxed, so a scheduled run that fails after the upgrade may be reporting a setting that was already broken. ## 2. Anthropic opens a private beta of Claude Managed Agents hosted entirely by the customer **[Operate Claude Managed Agents inside your security perimeter](https://claude.com/resources/articles/self-hosted-claude-managed-agents)** · Anthropic · claude.com, November 12, 2026 Managed Agents is Anthropic's set of APIs for building and deploying autonomous agents. It supplies "the agent harness, sandboxes, credential handling, and session history". It launched as a service Anthropic hosts, and Anthropic later let customers host the sandbox. In the new version, "every part of the service runs in your own cloud or on-premise". The customer operates the Managed Agents server, its database and the workers where agents run their shell and file tools. Model requests go where the customer chooses: Anthropic directly, Amazon Bedrock, Google's Agent Platform, Microsoft Foundry, or the customer's own gateway. Session history is stored in the customer's database. The post gives four design principles and says they apply in all three hosting arrangements. The harness that calls Claude is kept separate from the sandbox where the agent's code runs, and the agent loop, session history and vault secrets stay outside the sandbox. For MCP tools and supported API calls, a proxy outside the sandbox adds the credential to the request. Each environment can be limited to a list of allowed hosts. Every session leaves a record of each step, which exports over OpenTelemetry. Anthropic's stated reason: "you can't rely on a person checking every step, so the limits have to be built into the system itself." Access is by request through an Anthropic account team. The post gives no pricing, no system requirements and no customer results. We did not open the documentation it links to. **Why it matters:** For teams that wanted the whole agent service inside their own network, the post says, "Until now, that meant building the platform themselves." Those teams now have a vendor-built option to compare with their own harness, though only in private beta. The four principles also work as questions to put to a home-built setup: where the agent loop runs, where secrets live, what the sandbox can reach, and what record a run leaves. --- ## Also this cycle - **[Your Software Factory: Dream Factory or Treadmill?](https://tessl.io/blog/your-software-factory-dream-factory-or-treadmill)** · Patrick Debois, Tessl · November 11, 2026 · Debois sorts a factory's loops by what each run leaves behind. A treadmill loop gets today's work done, a learning loop makes the next round easier, and a dream loop tries something the team keeps postponing. The piece is an argument built on other vendors' write-ups, with no measurements of its own. His test for adding a loop: "ask what will be different after it runs."