Agent Intelligence: Hierarchical Planning (Plan-and-Act)

[samugit83]

Description

Replace the flat todo list with a two-level Planner → Executor hierarchy. The Planner sets strategic milestones with success criteria, iteration budgets, and contingency plans. The Executor (existing ReAct loop) works within the current milestone's scope. The Planner re-evaluates after each milestone completes or when the Executor gets stuck.

Why this matters

The agent's #1 failure mode on long engagements is losing strategic direction. After 15+ iterations, the flat todo list becomes stale and the agent drifts — repeating scans, trying random exploits, or spending 30 iterations on recon when it should have started exploitation 20 iterations ago.

A Planner solves this by:

1. Budget enforcement: each milestone gets a max iteration count. "Reconnaissance" gets 5 iterations — if it's not done by then, the planner forces a decision: proceed with what you have or ask the user.
2. Contingency planning: "If SQL injection fails on the web app, pivot to SSH brute force." Without this, the agent perseverates on a failing strategy because it has no pre-planned alternative.
3. Progress tracking with success criteria: "Reconnaissance is complete when we have open ports, services, and at least 1 potential vulnerability." The current todo list has no completion criteria — the agent never knows when to stop a phase.
4. Replanning on failure: when 3+ iterations pass with no progress, the planner activates a contingency plan automatically instead of looping.

Architecture (from README.EMPOWER.md Section 3.1)

                    ┌─────────────────┐
                    │     PLANNER     │  (Strategic: sets milestones, tracks progress)
                    │  "What to do"   │
                    └────────┬────────┘
                             │ Plan / Replan
                    ┌────────▼────────┐
                    │    EXECUTOR     │  (Tactical: existing ReAct loop)
                    │  "How to do it" │
                    └────────┬────────┘
                             │ Status update
                    ┌────────▼────────┐
                    │    PLANNER      │  (Evaluate: are we on track?)
                    └─────────────────┘

What already exists

• AgentState.todo_list — lightweight LLM-managed todo (no budgets, no completion criteria)
• Phase system (informational → exploitation → post_exploitation) — coarse-grained
• Basic failure loop detection (3+ consecutive failures triggers warning)
• conversation_objectives tracking in state

What needs to be built

• StrategicPlan Pydantic model with milestones, success criteria, iteration budgets, contingencies
• _planner_check conditional node before think — evaluates progress, triggers replan
• Replanning logic: budget exceeded → activate contingency or ask user
• Upgrade todo_list to strategic_plan in AgentState
• Planner LLM prompt that generates milestone-based plans from user objectives
• Progress evaluation: compare current state against milestone success criteria
• Frontend display of current plan/milestone status (optional)