Skip to main content
Ultron has a specific architecture. Understanding these concepts makes everything else in the documentation make sense.
A persistent, specialized AI worker. Ultron has five: Cortex (research), Specter (lead gen), Striker (sales), Pulse (content), Sentinel (infrastructure).Agents have domain expertise, access to specific tools, and memory of past runs. They coordinate with each other — Specter finds a qualified lead and creates a task, Striker picks it up and writes the email.
An on-demand specialized sub-agent. Invoked from chat with a specific objective. Each skill has its own model, tool set, token budget, and instructions.Skills run as isolated agents. You don’t configure them — you describe what you need and Ultron routes to the right skill. Results are saved to memory automatically after every run.Example skills: cold-email, competitive-analysis, linkedin-post, meeting-prep.
A unit of work assigned to a specific agent. Has a title, priority (P0/P1/P2), and status (inbox / in_progress / review / done / blocked).Tasks are how agents coordinate. Specter creates a task for Striker. Striker completes it and closes it. You can assign tasks directly from chat.
A visual artifact rendered inline in the chat thread. 30+ templates across five categories: Sales, Research, Content, Ops, and Data.Canvases are built with real data — Ultron runs web searches before generating them. They export as PDF or image. They render as React components inside the conversation.Examples: Sales Blueprint, SWOT Grid, Pricing Cards, Funnel Chart, Content Calendar.
A cross-session knowledge store. Four types:
  • user — your role, background, preferences
  • feedback — corrections and confirmations (both matter equally)
  • project — non-code business context (launch dates, team changes, priorities)
  • reference — external pointers (“See Linear board INGEST for context”)
Memory persists across sessions. Ultron selects the 5 most relevant entries per turn via a separate Sonnet side-query.
Automatic context management that keeps long sessions running without losing important information. Three layers:
  1. MicroCompact — removes stale tool outputs. No API call.
  2. Session Memory Compact — uses stored memories as compression baseline. No API call.
  3. API Digest — full LLM summary via fast model. Last resort.
Compression triggers automatically. You never need to manage it manually.
The standard protocol for connecting AI to external tools and data sources. JSON-RPC 2.0 over HTTP.Every Ultron session starts with 50+ MCP servers pre-configured: Brave Search, Puppeteer, GitHub, Apify, Tavily, Filesystem, Memory, and all your connected integrations.You don’t configure MCP servers manually — they load automatically based on your connected integrations.
A single continuous conversation. Transcript is persisted as JSONL after every turn. Sessions can be resumed exactly where you left off.Token usage is tracked per session. You can see cost breakdown in Settings → Usage.
The isolated cloud execution environment where complex tasks run. 50-minute runtime window. 50+ MCP servers available. Full filesystem and browser access.Sandbox sessions handle tasks that need more time or more tools than a standard API call allows. Available on Growth and Scale plans.
The integration layer that gives Ultron access to 250+ external services. Handles OAuth, token refresh, API versioning, and rate limit backoff for every connected tool.You connect a service once in settings. Composio handles everything else.