November 2 agent guide: Claude message threads and GLM-Legal

November 2, 2026 is an agent latency and domain-agent day. Anthropic opened message threads in public beta on the Messages API so long-running agents stop re-uploading the entire conversation every turn. Z.ai published GLM-5.3-FlashX-Legal for legal research, contract review, and drafting inside MCP-connected agent setups. xAI’s scheduled retirement of grok-imagine-image-quality also took effect. Bacancy Technology treats these as separate lanes: API transport, vertical agent model, and image API migration.

What shipped on November 2

Claude message threads for long agent loops

Without threads, every Messages API turn resends the full history, system prompt, tools, and tool results. Once a coding or computer-use agent carries hundreds of thousands of tokens, that upload becomes a large share of time to first token. With threads, the first request creates a server-side conversation. Later turns send only new messages plus the previous response id. Anthropic reports about a 35% drop in median time to first token for Claude Code requests between 500K and 1M tokens when threads are on.

Integration rules that matter

• Send the anthropic-beta: message-threads-2026-08-12 header on every request in the thread.
• Create with thread.type create and the full conversation. Continue with thread.type continue and previous_message_id.
• Do not resend system or tools on continue requests. Changing fingerprint fields such as model or tools returns thread_fingerprint_mismatch.
• Token usage is unchanged. Budget for the same spend with lower latency.
• Availability covers the Claude API, Claude Platform on AWS, Claude Models on Google Cloud (Anthropic managed), and Claude in Microsoft Foundry.

GLM-5.3-FlashX-Legal for MCP legal agents

Z.ai’s GLM-Legal line targets professional legal workflows: document understanding, legal reasoning, and multi-step execution such as research, contract review, cross-document checks, and drafting. The FlashX Legal API variant keeps a 1M-token context, mandatory thinking, function calling, structured output, and streaming. Z.ai recommends running it in an agent environment connected to specialized MCP services for laws, cases, company data, and Word document editing.

On Harvey LAB, Z.ai reports a 19.2% all-pass rate, 12.5 points above the base model, with independent evaluation notes pointing to Vals. LegalBench and PLawBench results are presented as parity with stronger closed models on a leaner footprint. Bacancy Technology’s reading for buyers: treat this as a domain agent candidate for controlled review desks, not as unsupervised legal advice.

xAI image slug retirement

Effective November 2, grok-imagine-image-quality is retired. Requests on image generations and edits are served by grok-imagine-image-2.0 with quality set to low. Shapes stay compatible, and grok-imagine-image (1.0) is untouched. Teams that care about quality or cost should call grok-imagine-image-2.0 and set quality to low, medium, or auto instead of relying on the redirect.

Pilot steps

1. 01
   Turn on message threads for one long agent

   Pick a Claude Code or Messages API loop that regularly exceeds a few hundred thousand tokens. Enable the beta header, keep a local transcript for replay, and measure TTFT before and after.
2. 02
   Bench GLM-Legal on a sealed contract pack

   Use non-production documents, require citation of source clauses, and route edits through an MCP document tool with human approval.
3. 03
   Update xAI image clients

   Replace grok-imagine-image-quality with grok-imagine-image-2.0 and pin quality. Confirm response.model in logs after November 2.

Conclusion

November 2 is about making long agents faster to first token, giving legal teams a FlashX-class domain model with MCP hooks, and forcing an explicit xAI image migration. Latency wins do not change token economics. Domain models still need human review on consequential work.

Related reading

• November 1 Mage-VL-Cowork: /blog/november-01-mage-vl-cowork-computer-use-2026
• October 31 Super-agent and Claude policy: /blog/october-31-deepseek-superagent-claude-code-2026
• MCP topic map: /topics/mcp
• Agentic AI topic map: /topics/agentic-ai

Need message-thread and MCP legal pilots on a production stack?

Bacancy Technology helps engineering and legal-ops teams instrument agent latency, MCP tool allowlists, and human-approval gates. Consulting: /consulting/ . Contact: /contact/ .

Sources: Anthropic Platform release notes, Claude Platform docs . Message threads docs, Claude Platform docs . Message threads article, Claude blog . Z.ai GLM-Legal, z.ai — glm legal . Z.ai Flash docs, docs.z.ai — glm 5.3 flash . xAI migration guide, docs.x.ai — imagine image quality nov 2 .