Goal Engine

The Goal Engine enables WASP to execute complex, multi-step objectives asynchronously. Instead of processing a request in a single conversation turn, the Goal Engine creates persistent goals that run in the background.

Overview

User: "Monitor BTC price and alert me when it crosses $70k"
            │
            ▼
    GoalOrchestrator.create_goal()
            │
            ▼
    PlanGenerator → TaskGraph
    ┌─────────────────────────────┐
    │ Task 1: Subscribe to BTC    │
    │ Task 2: Check price loop    │
    │ Task 3: Send alert if >70k  │
    └─────────────────────────────┘
            │
            ▼ (GoalTickJob runs every 15s)
    Execute tasks → update state → complete

Key Components

GoalOrchestrator (src/goal_orchestrator/orchestrator.py)

The main coordinator. Responsibilities:

• Creates and persists goals in Redis (goals hash)
• Ticks active goals via GoalTickJob
• Manages up to 3 concurrent goals (configurable)
• Handles budget enforcement (tokens, replans, memory)
• Sorts goals by priority (user=8, agent=6, autonomous=3)

PlanGenerator (src/goal_orchestrator/planner.py)

Converts a natural language objective into a TaskGraph:

• Uses the LLM with a 4,000-token planning budget
• Generates a DAG of tasks with dependencies
• Skill list in prompt is capped at 1,200 characters (PLAN_SKILL_BUDGET)
• Output is a structured JSON TaskGraph

TaskGraph

The execution plan — a directed acyclic graph of tasks:

TaskGraph(
    tasks=[
        Task(id="t1", description="Search for current BTC price", skill="web_search"),
        Task(id="t2", description="Check if price > threshold", depends_on=["t1"]),
        Task(id="t3", description="Send Telegram alert", depends_on=["t2"]),
    ],
    parallel_groups=[["t1"]],  # Tasks that can run in parallel
)

PlanCritic (src/goal_orchestrator/plan_validator.py)

A second LLM pass validates generated plans before execution:

• Checks for logical gaps, missing error handling, impossible tasks
• Can reject the plan and trigger regeneration
• Uses 1,200 tokens (fast, lightweight validation)
• Enabled via PLAN_CRITIC_ENABLED=true (default)

Goal States

PENDING → PLANNING → ACTIVE → COMPLETED
                   ↘          ↗
                    PAUSED → ACTIVE (auto-resume)
                   ↘
                    FAILED

State	Description
PENDING	Goal created, not yet started
PLANNING	PlanGenerator running
ACTIVE	Tasks executing
PAUSED	Backpressure or waiting (auto-resumes)
COMPLETED	All tasks done, summary sent
FAILED	Max replans or budget exceeded

Budget Controls

The Goal Engine enforces strict budgets to prevent runaway execution:

Setting	Default	Description
GOAL_BUDGET_MAX_TOKENS_PLANNING	4,000	Token limit for plan generation
GOAL_BUDGET_MAX_TOKENS_EXECUTION	20,000	Total execution token budget
GOAL_BUDGET_MAX_REPLANS	5	Max plan regenerations per goal
GOAL_BUDGET_MAX_MEMORY_BYTES	1 MiB	Working memory limit per goal
GOAL_DEFAULT_MAX_STEPS	50	Max task steps before forced completion
GOAL_DEFAULT_MAX_RUNTIME	3600s	1-hour hard timeout

Priority System

Goals are prioritized 1-10 and sorted before each tick:

Source	Priority	Description
User request	8	Direct user message
Agent-created	6	From sub-agent
Autonomous	3	From AutonomousGoalGeneratorJob
Default	5	Unspecified source

Autonomy Modes

Mode	Description
full	Execute all tasks without asking (default in Sovereign Mode)
semi	Confirm destructive or external actions
assist	Request approval for every action

Completion Synthesis

When a goal completes, WASP sends a rich Telegram message:

• Summary of what was accomplished
• Key outputs from tasks
• Time taken and resources used

Example: Creating a Goal via API

# Via dashboard API
curl -X POST https://agentwasp.com/api/goals \
  -H "Content-Type: application/json" \
  -d '{
    "objective": "Monitor ETH price hourly for 24 hours, log to memory",
    "priority": 8,
    "source": "user"
  }'

Replan Loop

If a plan fails mid-execution, the Goal Engine can replan:

1. Analyzes what failed and why
2. Generates a new TaskGraph accounting for partial progress
3. Resumes from the last successful task
4. Limited to GOAL_BUDGET_MAX_REPLANS (default: 5)

Stability Protections

To prevent thrashing:

• REPLAN_STORM_COUNT = 6 — max replans in a 10-minute window
• CONSECUTIVE_FAILURE_LIMIT = 5 — fail goal after 5 consecutive failures
• PAUSED goals auto-resume after backoff, fail after 10 minutes paused