Skip to main content

WASP Documentation

Current version: v2.2 — See what's new in v2.2

WASP is a fully autonomous AI agent platform. It receives natural language instructions via Telegram or its web dashboard, decomposes them into executable plans, runs those plans using a rich skill library, and continuously learns from its experience.

What makes WASP different​

Unlike simple chatbots or prompt-chaining tools, WASP operates as a self-governing agent:

  • Plans autonomously — given any objective, WASP generates a multi-step execution plan via a dual-layer planner with a built-in critic
  • Executes with real skills — web search, full browser control (Chromium), Python execution, shell access, Gmail, file operations, and 37 built-in skills total
  • Learns from experience — behavioral corrections, episodic memory, procedural memory, skill evolution, and goal-level self-reflection
  • Reasons about time — a temporal world model tracks real-world changes and trends across conversations
  • Runs in the background — 27 scheduled jobs operate continuously without user interaction, including response validation, dream consolidation, autonomous goal generation, background perception, and opportunity detection
  • Self-governs resources — per-user rate limiting, concurrent goal caps, and LLM budget enforcement via the Resource Governor

Platform Architecture​

Key Systems​

SystemDescription
Goal EngineDecomposes objectives into TaskGraphs, executes with plan critic validation
Skill System37 built-in skills across 5 capability levels, supports custom Python skills
Memory18 persistent memory systems (11 primary + 7 auxiliary): episodic, semantic, procedural, visual, vector, KG, self-model, temporal, goal-scoped, ranked retrieval, reflection memory, behavioral rules, learning examples, dream log, recovery memory, skill patterns, entity states, and state predictions
Scheduler27 background jobs covering health, learning, perception, opportunity detection, and autonomous goal generation
Resource GovernorPer-user rate limiting: goal slots, LLM budget, API call caps with safe degradation
Decision LayerPre-LLM heuristic classifier with 13 fast-paths, routes requests to 5 strategies before LLM is invoked
Multi-AgentSpawn and coordinate multiple sub-agents with independent goal queues
Integrations40+ connector types: Slack, Discord, GitHub, Notion, Telegram, Gmail, smart home, and more

Version 2.2 — March 2026​

New in v2.2:

  • deep_scraper Built-in — Playwright/Crawlee containerized scraper migrated to permanent built-in. YouTube transcript extraction via network interception, JS-heavy/anti-bot page rendering. SSRF protection with DNS-level IP validation (RFC1918/loopback/link-local/reserved blocking, fail-closed).
  • 37 built-in skills — Up from 27. Added: deep_scraper, browser_screenshot_full_page, browser_smart_navigate, browser_deep_scrape, list_reminders, delete_reminder, google_calendar, meta_orchestrate, agent_manager, and more.
  • Dashboard streaming — SSE-based live progress: POST /chat/stream endpoint shows each LLM round and skill call in real-time. sessionStorage chat history persists within tab session (30-message cap).
  • Capability map hardened — All 37 skills explicitly mapped in _CAPABILITY_MAP. No skill defaults to untracked CONTROLLED level.
  • Response Validator extended — deep_scraper added to _PRICE_GROUNDING_SKILLS for price hallucination detection.
  • Auto-detect routing fixed — YouTube URL auto-detection now routes to deep_scraper (MONITORED) instead of shell (RESTRICTED), enforcing proper capability enforcement.
  • Production stabilization — Replan storm detection tightened (REPLAN_STORM_COUNT=3, window=5min). Goal budget max replans=3. Storm marks goal FAILED with partial output.
  • Pre-Production audit — 21 bugs fixed across autonomous.py, handlers.py, executor.py, orchestrator.py, temporal.py, learning.py, security (SSRF, path traversal, CSRF binding, media exposure).

Version 2.1 — March 2026​

New in v2.1:

  • Multi-Agent System — Full AgentOrchestrator with AgentRuntime, CapabilitySandbox, inter-agent message bus (PostgreSQL-backed), global token budget (Redis per-minute), CPU backpressure
  • AgentManagerSkill — LLM can create/list/pause/resume/archive agents at runtime
  • Goal priority axis — Goal.priority (1-10) + Goal.source; user goals=8, agent goals=6, autonomous=3
  • Self-Integrity Monitor — Cross-checks self-model vs actual skill rates every 6h; reports to dashboard
  • Cognitive Pressure Index — Composite 0-100 metric (goals/errors/latency/memory/CPU); actuators skip when CPI > 80
  • Epistemic calibration fixed — Symmetric ±0.015 penalty (was asymmetric)
  • Self-Model durability — File backup at /data/memory/self_model.json; Redis miss falls back to file
  • Circuit Breaker persistence — Redis state save on every transition; restored on first call

Quick Navigation​