AI Agent
Natural language workflow generation and intelligent assistant
AI Agent
B3OS includes a built-in AI agent that turns natural language into fully functional workflows. Instead of manually wiring nodes together, describe what you want and the agent builds it for you.
What It Does
The agent operates in three modes:
How It Works
You describe what you want
Type a message like "Send me a Slack alert whenever ETH drops below $2,000" or "Create a weekly portfolio rebalancer for my Morpho positions."
The agent plans and builds
The agent searches available actions and triggers, resolves schemas, fetches any needed on-chain data, and assembles a valid workflow definition.
You review and deploy
The generated workflow appears in the canvas editor. Review the nodes, tweak any parameters, and publish when ready.
17+ Specialized Tools
The agent has access to real-time data and platform knowledge to build accurate workflows.
| Tool | What it does | |------|-------------| | Action search | Browse 280+ actions across 21 providers | | Schema resolution | Look up input/output schemas for any action | | ABI fetcher | Fetch on-chain contract ABIs for EVM interactions | | Template browser | Search pre-built workflow templates | | Trigger lookup | Find the right trigger type for any use case |
| Tool | What it does | |------|-------------| | Token prices | Live prices from CoinGecko and exchanges | | Wallet balances | Check balances across chains | | DeFi positions | Current lending/LP positions and yields | | Yield optimization | Compare APYs across protocols | | Prediction markets | Polymarket odds and orderbook data |
| Tool | What it does | |------|-------------| | Org documents | Semantic search across uploaded docs | | Google Sheets | Read spreadsheet data for context | | Web search | Look up current information | | Analytics queries | Discover and run Dune queries | | Existing workflows | Reference your org's workflow patterns |
Knowledge Sources
The agent uses a 3-layer retrieval system to generate accurate, context-aware workflows:
Layer 1: AI Profile
Static system context including B3OS feature documentation, action/trigger specifications, and platform capabilities.
Layer 2: Org Knowledge
Semantic search over your organization's uploaded documents and notes using pgvector embeddings. The agent understands your team's specific context.
Layer 3: Recipe Index
Fuzzy matching against a library of workflow patterns and common automation recipes. Helps the agent suggest proven configurations.
Model Architecture
The agent routes to different models depending on the task:
| Task | Model | Why |
|---|---|---|
| Workflow generation and tool-calling | GPT via OpenRouter | Best tool-calling and structured output |
| Conversation memory extraction | Claude Haiku | Fast, accurate summarization |
| Cost-optimized fallback | Nvidia Nemotron | Free-tier backup for simple queries |
| Vision (screenshot analysis) | Claude Sonnet | Strong image understanding |
Streaming and Sessions
Responses stream in real-time via SSE (Server-Sent Events). You see the agent's thinking and output as it generates, not after.
Long conversations are automatically summarized to preserve key context without hitting token limits. The agent remembers what you discussed earlier in the session.
Chat sessions persist across page reloads. You can pick up where you left off or start a new session for a different task.
Attach screenshots of dashboards, charts, or interfaces. The agent processes images and incorporates visual context into workflow generation.