[Feature] Multi-step task execution (agentic workflows)

[Wan-ZL]

Description

Transform Genesis from a tool-calling chatbot into a true agent that can plan and execute multi-step tasks autonomously. This is the key capability gap between Genesis and the 2026 AI agent landscape.

Current State

Genesis has individual tools (web_fetch, shell, calendar, repository, etc.) but they operate in isolation. The LLM calls one tool at a time in a single request-response cycle. There is no ability to:

• Plan a sequence of actions
• Execute multiple steps toward a goal
• Handle errors mid-execution and retry/adjust
• Report progress on long-running tasks
• Chain tool outputs as inputs to subsequent tools

Target State

Users should be able to say things like:

• "Research the top 5 restaurants near me, check their reviews, and create a comparison table"
• "Back up my important files, compress them, and upload to my server"
• "Find all Python files that import deprecated modules and suggest replacements"

Genesis should decompose this into steps, execute them, handle failures, and deliver results.

Acceptance Criteria

• Task planning: LLM decomposes user request into ordered steps
• Step execution: Each step can invoke one or more tools
• Progress reporting: User sees real-time progress updates
• Error handling: Failed steps can be retried or skipped with explanation
• Tool chaining: Output of one tool can be input to another
• Task history: Completed tasks stored for reference
• Cancellation: User can cancel a running task
• Permission checks: Each step validates permissions before execution
• Tests: Unit tests for task planning and execution
• Streaming: Progress updates streamed to frontend via SSE

User Research

• "Agentic AI" is the defining trend of 2026 (Microsoft, MIT Technology Review, TechCrunch)
• Users want AI that DOES things, not just TALKS about things
• OpenClaw's popularity (171k stars) proves demand for autonomous task execution
• MCP protocol enables multi-agent workflows where agents chain actions
• Organizations report 40-60% efficiency gains with agentic workflows

Priority Rationale

HIGH for Phase 9. This differentiates Genesis from chatbots. Combined with MCP support (#51), this creates a genuine agent platform. However, it depends on the existing tool system being stable and MCP integration providing more tools to chain.

Technical Notes

• Use ReAct (Reason + Act) pattern for task decomposition
• Leverage existing tool registry for execution
• SSE events for progress streaming (existing infrastructure)
• SQLite for task state persistence
• Consider a "plan review" step where user approves the plan before execution