Shipped foundations plus where agent is heading in 2026. Planned items are aspirational, not yet implemented.
- Two-loop core: follow-ups outer, tools + steering inner
- Parallel and sequential tool execution
- Steering, follow-up queues, cooperative cancellation
- Ten lifecycle events for UI observation
- AgentTool model + declarative ToolDef / @action / Param
- Built-ins: filesystem, bash, current_time, submit_result
- before/after tool hooks, context transforms, arg validation
- Tool registry SPI and shared tool_ctx sink
- First-class sub-agents a tool can spawn and await
- Structured hand-off between specialised agents
- Budgets and depth limits for delegated work
- MCP (Model Context Protocol) tool servers as AgentTools
- Pluggable long-term memory hooks around the loop
- Replayable transcripts for eval harnesses