Agent tool orchestration is transitioning away from static JSON schema expansion and granular API calls toward harness-side code execution, sandboxing, and progressive context disclosure.
Evidence
- ‘Token Optimization for Agents via Progressive Disclosure and Code Mode’ and ‘What Is Codemode’ demonstrate that executing composed tool calls in an isolated harness-side code environment prevents context exhaustion and intermediate output bloat.
- ‘Migrating a Personal OS’ and ‘Agent Harness Study’ reveal that replacing granular tool APIs with server-side sandboxed Python execution and trimming tool schemas cuts inference spend by up to 3x while substantially lowering latency.
- ‘Curating Always-Loaded Context’ and ReFold empirically prove that continually appending static rules or raw turn histories degrades model instruction compliance and invalidates prefix cache reuse.
Implications
- Agent tool frameworks will increasingly standardize on programmatic execution environments (code mode and sandboxes) rather than cataloging individual REST/JSON tool schemas in prompt space.
- Agent runtime design is decoupling orchestration prompts from execution runtimes to curb token overhead and maintain prompt cache efficiency during long-horizon tasks.
Concepts
Agents Tool Use Context Engineering MCP Code Agents
Confidence
high