AI
Tool Routing Now Lives in the Model
Before this year, developers built string-matching routers (“if the model mentions SQL, give it the database tool”) and wrapped them in retry decorators because tool calls failed often enough to need backoff. Those routers were brittle. Adding a new tool typically broke them. Now Claude selects from its full tool set directly. Tool selection accuracy is high enough that pre-filtering makes performance worse. When a tool errors, Claude sees it, recovers, and retries without harness involvement.
Context Management Without Custom Scaffolding
Long-running agents used to require custom memory: chunking, RAG, summary calls after N tokens, manual cache breakpoint management. Anthropic now offers 1 million token context at flat pricing, paired with server-side compaction and context editing. That collapses most of the scaffolding to a few lines of config. A practical technique: clear tool results at the end of every turn. Screenshots, search results, and file reads can be pruned while keeping the decisions Claude made from them, saving tokens in real time.
Code Execution as a Built-in Sandbox
Previously the write-run-fix loop was entirely developer work: find a VM provider, spin up a sandbox, push code to it, parse tracebacks, feed errors back into the model, repeat. Anthropic now ships a code execution tool that gives Claude a hosted sandbox server-side. The entire iteration loop runs inside a single API turn. No harness round trips. Claude uses the sandbox for stateless compute, data analysis, and custom libraries without touching the local file system, then reaches back to local bash when it needs repo access or local context.
Computer Use at Native Resolution
Computer use previously required pixel scaling code: downscale a 1080p screenshot to fit Claude’s pixel limits, track the scale factor, multiply click coordinates back up. Opus 4.7 accepts native screenshots and returns one-to-one pixel coordinates up to 1440p. OS World scores jumped from below 50% to 78% in under 12 months. Gonzalez showed this in a live demo where Claude found and fixed two bugs in a project management dashboard by testing the app in Chrome itself, including drag-and-drop. The closing argument: scaffolding written to compensate for model unreliability has a half-life of months. The code worth writing connects Claude to data only you control.
Notable Quotes
build last year actually ships with the model today. Lucas Gonzalez · ▶ 1:02
usually makes things worse, not better. Lucas Gonzalez · ▶ 4:17
we just reported 78% on Opus 47. Lucas Gonzalez · ▶ 14:32
Key Takeaways
- Tool routers built from heuristics are obsolete; Claude selects and retries tools on its own.
- Server-side compaction and 1M-token flat pricing replace most custom memory systems.
- The code worth investing in connects Claude to your proprietary data, not to model workarounds.