Agent Execution: Runtime, Modes & Caching

**Value proposition**

As agents take on longer, more complex tasks, the speed, cost, and flexibility of execution becomes a limiting factor. Elastic is rebuilding the execution harness that powers Agent Builder, introducing distinct agent execution modes, introducing code based tools and improved prompt caching to reduce token usage, and multi-model routing for optimal performance.

**Expected outcome**

- A new execution runtime that introduces a virtual file system and bash-like tool for the agent to offload and process context more effectively
- An agent planning mode ("Plan") that sequences tool/skill use into a visible multi-step plan before executing, and lays the groundwork for additional modes later.
- Prompt caching that reduces both latency and inference cost for repeated or similar prompts.
- Greater control over tool execution including ability to configure execution timeouts per tool

**Key user stories**

- As a user, I want the agent to manage context efficiency so I can allow it to perform more complex tasks without incurring significant higher costs
- As a user with a complex request, I want the agent to show me its plan before it starts calling tools, so I can assess and adjust the plan early and ensure I am getting a relevant response.
- As a platform admin, I want lower inference costs at high-volume usage so Agent Builder scales affordably across my org.

**Phasing**

- Phase 1 (Now): Execution runtime improvements. Prompt caching, tool timeouts
- Phase 2 (Next): Agent Modes: Planning