The agent-native
runtime
Durable execution. Native MCP + A2A. Full observability.
Built in Rust. Authored in Python.
Production agents have
unsolved problems
Agents crash and lose everything
Durable graph execution with event sourcing
Agents are locked in silos
Native MCP + A2A protocols
You can't see what's happening
OTel GenAI traces + checkpoint replay
Autonomous agents run forever
Compile-time autonomy constraints
Human approval is bolted on
Human-in-the-loop as a native node type
Python orchestration doesn't scale
Rust async scheduler, microsecond overhead
Python you
already know
Decorate functions with @task and @tool. The runtime handles durability, cost limits, and telemetry.
from jamjet import task, tool @tool async def web_search(query: str) -> str: """Search the web for current information.""" ... @task(model="claude-sonnet-4-6", tools=[web_search]) async def research(question: str) -> str: """Search first, then summarize clearly.""" result = await research("Latest AI agent trends?") print(result)
Everything production demands
Durable Graph Execution
Every step checkpointed with event sourcing. Crash the process, restart it — execution resumes exactly where it stopped.
Native MCP + A2A
Connect to any tool server. Expose your tools. Delegate to external agents. Both protocols built in, not bolted on.
Full Observability
OpenTelemetry GenAI traces. Per-node cost attribution. Checkpoint replay. Know exactly what happened and why.
Human-in-the-Loop
Native workflow node for approvals. Durable suspension — the process sleeps until a human responds, even across restarts.
Autonomy Limits
Token budgets, cost caps, iteration limits. Enforced by the runtime, not by convention. Agents can't run away.
Native Eval Harness
LLM-as-judge, assertion, latency, and cost scorers. Run evals as workflow nodes. CI exit codes on regression.
Typed Schemas
Pydantic and JSON Schema validation at every step boundary. No loose dicts. Catch errors before they propagate.
Rust Core
Async scheduler built in Rust. Microsecond dispatch overhead. Python is for authoring, not for orchestration.
Multi-Model
Any OpenAI-compatible API. Anthropic, Ollama, Groq, Azure — swap models per task with a single config change.
Your next paper,
reproducible by default
Cartesian product over models, prompts, and temperatures. Parallel execution. Results export to CSV, LaTeX, and JSON with one call.
ReAct, plan-and-execute, critic, reflection, consensus, and debate — compiled to IR sub-DAGs. Compare strategies across the same dataset.
Every decision checkpointed. Replay any execution. Fork with modified inputs. Perfect for ablation studies and failure analysis.
Start building
Get a durable workflow running locally in under 10 minutes.