Guides¶
These guides take a system from its first execution profile to production. Start with an agent, choose how work is orchestrated, connect only the capabilities you need, then add serving, testing, and operations.
Choose the execution path first¶
| Need | Start here |
|---|---|
| One autonomous agent reasons through a task | AutoAgent |
| Application code owns every execution step | CustomAgent |
| Your application already knows the workflow DAG | Workflow DAGs |
| Peers must derive and claim work from a source goal | Durable Goal Execution |
Durable Goal Execution is an independent guide because it has a different state contract from a declared workflow: immutable attempts, source obligations, revision-fenced reconciliation, and separate execution, obligation, and response status. Learn Workflow DAGs first if you are new to multi-step execution.
1. Build the agents¶
-
AutoAgent
The framework brings the brain. Describe the task; the agent plans, generates code in a sandbox, and self-repairs. Use when you can describe the work but not code it.
-
CustomAgent
You bring the brain. Plain Python in
execute_task, deterministic control, same mesh. Use when you can code the logic. -
System Prompts
Define an AutoAgent's role, output contract, verification behavior, and failure handling.
-
Custom Sub-agents
Extend the Kernel with a specialized execution unit when built-in roles are not enough.
-
Adapters
Bring existing Python, LangChain, or CrewAI objects into the Mesh without a full rewrite.
2. Orchestrate work¶
-
Workflow DAGs
Declare known steps, roles, dependencies, parallel branches, and output chaining in application code.
-
Durable Goal Execution
Compile an unknown source goal into capability-addressed work that peers claim independently, with durable obligation truth and selective repair.
-
Human-in-the-Loop
Pause only at genuine human boundaries, then resume the same durable work.
Multiplayer: teams and agents working together¶
-
Build a Multiplayer Agent
From nothing to a team using it in Slack: the agents that do the work, a shared room, the team on the mesh, and a room agent people talk to.
-
Rooms
Shared files that agents propose changes to and people review, with merging, roles and a permanent record of every decision.
-
People as Peers
Put people on the mesh with capabilities, route work to them, and let them answer from wherever they are.
-
Chat Surfaces
Serve a room in Slack through Nexus, or in your own product through one small interface.
3. Connect capabilities¶
-
Integrations
Discover the typed atom catalog for business systems, storage, developer tools, communication, and provider readback.
-
Nexus Credentials
Register provider connections while keeping raw credentials outside agent reasoning and generated code.
-
Knowledge Base
Ingest internal documents and add local retrieval to research work.
-
Internet Search
Configure grounded and multi-provider search with ranking and resilience.
-
Browser Automation
Add Playwright-backed interaction for websites that require a real browser.
4. Serve, test, and operate¶
-
FastAPI
Embed agent lifecycle into a FastAPI application.
-
Chat Endpoint
Expose synchronous and streaming conversational interfaces.
-
Observability
Inspect traces, events, costs, workflow state, and Prometheus metrics.
-
Testing Agents
Exercise agent behavior with deterministic Mesh, peer, and LLM mocks.
-
Testing Atoms
Validate integration functions structurally and against connected providers.
-
AI Editor Setup
Teach Copilot, Claude Code, Cursor, and other coding agents the current JarvisCore contracts.
-
Production Deployment
Harden storage, credentials, sandboxing, networking, budgets, and monitoring.