October 31 agent guide: DeepSeek Super-agent and Claude Code policy

October 31, 2026 closed the window with architecture research and more Claude Code policy. DeepSeek published deepseek-ai/dsh-superagent-prototype and stated it is not an official product. Claude Code added layered managed policy, pooled budgets, an ultrathink disable flag, Bedrock input tagging, and a cloud MCP form card. Write blocks and gateway groups from October 30 stay in that day’s guide.

What shipped on October 31

DeepSeek Super-agent is architecture research

On October 31 DeepSeek published deepseek-ai/dsh-superagent-prototype on GitHub. The README describes a research prototype and states that it is not an official DeepSeek product. The design splits a Router from isolated Workers. The Router has no Bash. Workers do not talk to humans. Admission uses tickets and mandates. That isolation story is the part worth reading for teams that are drawing multi-agent graphs.

The repository is useful as a pattern, not as a platform. Prompt injection and credential handling are still listed as open problems in the prototype framing. A later design essay dated November 1 points at the same prototype. It is not a second product launch, so November 1 does not get its own shipping brief.

Claude Code layered policy and pooled budgets

October 31 adds CLAUDE_CODE_DISABLE_ULTRATHINK, pooled group budgets through pooled_budget, managed.policy_mode set to layered, Bedrock guardrail.input_tagging, and a cloud MCP form card. Pooled budgets are how a team stops one long agent from consuming a whole group’s quota. Layered policy is how an enterprise keeps a shared fleet from each developer carrying a private exception list. Write blocks remain the October 30 setting: /blog/october-30-claude-code-write-blocks-2026.

Pilot steps

1. 01
   Turn on layered policy and pooled budgets

   Set managed.policy_mode to layered on shared Claude Code fleets. Attach pooled_budget so one long job cannot drain the group. Confirm Bedrock guardrail.input_tagging if the fleet is on Bedrock.
2. 02
   Read Super-agent as a graph review, not a deploy

   Ask whether the current planner has a shell, whether workers can speak to users, and whether there is an admission step before a tool runs. Do not clone the prototype into production.

Common mistakes

• Deploying DeepSeek Super-agent as an internal platform. The README calls it a research prototype, not an official product.
• Treating a November 1 design essay as a second Super-agent launch.
• Skipping October 30 write blocks because October 31 added layered policy. Both days are required.
• Leaving Opus 4.8 fast mode in place. Removal is still November 30 with no fallback.

Conclusion

October 31 is policy and a prototype. Layered Claude Code settings belong on shared fleets. Super-agent belongs in a design review, not in a sprint board. November 1 did not add a separate product launch.

Related reading

• October 30: Claude Code write blocks and gateway groups: /blog/october-30-claude-code-write-blocks-2026
• October 29: Cursor marketplace incident and Mistral Large 4: /blog/october-29-cursor-marketplace-mistral-large-4-2026
• October 28: Copilot MCP connectors and Gemini 4 Carbon: /blog/october-28-copilot-mcp-gemini-carbon-2026
• October 27 agent guide: /blog/october-27-agent-guide-copilot-sdk-lfm3-2026
• Agentic AI topic map: /topics/agentic-ai

Need this policy board applied to a production stack?

Bacancy Technology reviews Claude Code policy layers and multi-agent isolation patterns against the systems a team already runs. Speak with the consulting team or use the contact form. Consulting: /consulting/ . Contact: /contact/ .

Sources: DeepSeek Super-agent prototype, GitHub . Claude Code changelog, Claude blog .