Describe what you want. Smith generates the multi-agent graph, agents, tools, and routing, then deploys it live with tracing.
No SDK to learn · no infrastructure to manage · deploy in one click
One conversation. The whole system designed, deployed, and traced.
You describe the outcome. Smith handles topology, prompts, routing, and hosting, and you stay in control of every node.
Describe the system you want in plain language.
Smith generates the agent graph node by node, with routing welded in.
Tune any agent's model, temperature, tools, and system prompt.
One action compiles the graph and publishes a versioned live endpoint.
Watch real traffic flow through the graph with live traces and metrics.
Generation, graph, inspector, deploy, and traces, all in one place. No notebook, framework, and dashboard to juggle.
Describe the outcome. Smith turns it into a structured spec, picks a topology, drafts prompts, and assigns tools.
Every system is a legible directed graph of agents, tools, and routing. Not a black box.
Smith suggests a model per role. Change it, plus temperature, tools, and prompt, on any node.
Compile and publish to a versioned, regional endpoint in seconds.
Runs, p50 latency, and success rate, streamed as traffic arrives.
Every system is a structured spec you can read, edit, and export. Stay in the studio or take the graph with you.
{ "system": "support-copilot", "nodes": [ { "id":"supervisor", "role":"orchestrator", "model":"claude-opus-4.8" }, { "id":"billing", "role":"specialist", "tools":["stripe","kb"] } ], "edges": [ { "from":"supervisor", "to":"billing", "type":"conditional" } ], "deploy": { "region":"iad1", "endpoint":"smith.run/support-copilot" } }
Most tools stop at a single agent. Smith generates the orchestration: supervisors, specialists, parallel work, and feedback loops.
A router delegates each task to the right specialist and decides when the job is done.
Each step hands its output to the next, in a fixed order, start to finish.
Run independent agents at once, then aggregate their results into one answer.
An agent acts, a critic checks, and the work loops until it meets the bar.
Real descriptions, and the multi-agent systems Smith forged from each.
The graph is an output of a conversation, not something you assemble by hand. That makes the generated system legible, editable, runnable, and observable, all in one place.
The whole system is a graph you can see and understand, never a hidden chain.
Every agent, tool, and edge is inspectable and tunable before and after deploy.
One action takes a draft to a live, versioned endpoint on the Smith runtime.
Live traces and metrics are built in, right where you built the system.
Visual builders give you a blank canvas to wire agents by hand; code-first tools like LangGraph have you write the graph in code. Smith generates the multi-agent graph from your description, then lets you tune any node and deploy with tracing, in one place. The graph is an output, not a chore.
No. You describe the system in plain language and Smith generates the graph. The framework details are handled for you, though developers can read and edit the underlying spec.
The Smith runtime compiles each generated graph into an executable agent system, hosts a versioned endpoint, and streams traces. The studio is the design surface on top of it.
Yes. Select any agent to change its model, temperature, tools, and system prompt. The generated graph is a starting point you stay in control of.
Smith auto-assigns and suggests a model per role: stronger ones (Claude Opus) for reasoning and security, faster ones (Sonnet, Haiku) for mechanical steps. It's a suggestion, change the model on any node to fit your preference, budget, or latency.
Smith Studio is in early access. You can explore the full build-to-deploy loop in the live prototype now; real generation and runtime are rolling out in phases.
Describe the system you want, and watch Smith forge it into a deployable multi-agent graph.