Fetching from the wire…
Public story · 2026-07-25 · high
A new header lets you defer tools and add or drop them mid-conversation instead of rewriting the whole prompt array.
Why now: It ships under a beta header dated July 1, aimed squarely at multi-phase agent systems already hitting this exact cache wall.
Anthropic shipped a beta with Opus 5 that stops mid-conversation tool changes from wiping a conversation's entire prompt cache, per the mid-conversation system messages guide in its developer docs.
Agent systems that move through several phases used to pay a full cache rebuild every time they swapped a tool. The new beta lets that swap happen without the rebuild.
The tools array sits earlier in the hashed request prefix than the top-level system field, per the docs. Edit that array to add or drop a tool, and the cache for the whole conversation invalidates, not just the request.
The beta flips the order of operations. You declare the full tool set up front and mark the ones you don't need yet as defer_loading: true.
To bring a deferred tool in or take an active one out, you send a role: "system" message carrying a tool_addition or tool_removal block naming the tool, instead of editing the array again.
MCP tools get the same treatment. A single tool swaps via mcp_tool_reference; a whole server swaps via mcp_toolset_reference.
The docs don't say what a deferred tool costs while it sits dormant, or whether the mechanism will reach models older than Opus 5.
Each link below shares sources, entities, or timing with this story.
This one rearranged my week. An essay published August 4 walks through Databricks' independent benchmark of coding harnesses against its own multi-million-line codebase. Pi, a harness with four built-in tools and a system prompt under 1,000 tokens, paired with Opus 4.8 at xhig...
The June 15 release of Claude Code (v2.1.178) is small on the changelog and large in practice. Two changes matter for anyone running agentic workflows daily. First, permission rules now match on parameters, not just tool names. You can write Agent(model:opus) to block Opus sub...
The Compaction API enables server-side context summarization for effectively infinite conversations without context limits, available on Opus 4.6. MCP servers can now request structured input mid-task — interactive dialogs with text fields, dropdowns, and checkboxes — via new...
Three separate Anthropic changes over about two weeks point the same direction, and none of them announced themselves as a strategy. Claude Code 2.1.238 added claude self-hosted-runner --defer-shutdown-max-min, which keeps serving attached sessions on SIGTERM, parks whatever's...
August 6: Sapiom raised $35M for a router that sends each model call to the cheapest capable model. August 7: Databricks published its internal cost playbook. August 8: Toolport hit Product Hunt with a free MCP gateway cutting tool-definition overhead 96%. Three parties, zero...
A solo Claude Opus 4.5 agent spent $9 and 20 minutes building a retro game. It was broken. The same model, wrapped in Anthropic's multi-agent harness, spent $200 over 6 hours and produced a fully playable game with physics, sprite editors, and AI integration. Anthropic's engin...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.