Production-safe
AI agents.
JamJet is an open-source runtime that gives AI agents policy controls, audit trails, human approval, crash recovery, cost governance, and memory — without locking you into a cloud or framework.
Not another agent framework. The safety layer behind them.
An open-source governance runtime for agentic systems.
Six layers between your agent and production failure.
Other tools observe what agents do. JamJet actively prevents unsafe behavior, recovers from failures, and creates audit evidence — at the runtime layer.
Block unauthorized tool calls, restrict model usage, scope delegation permissions. 4-level policy hierarchy enforced before execution, not after.
Immutable, append-only log of every decision, tool call, delegation, and approval. Separate from execution data. Retention-aware. Ready for compliance export.
First-class pause/resume/approval nodes. Workflows survive restarts while waiting for human input. Not a callback — a durable primitive.
Event-sourced durable execution. Kill the process. Restart. It resumes from the exact checkpoint. Completed steps are never re-run.
Per-workflow and per-agent token and dollar budgets. Enforced by the runtime, not by application discipline. Automatic escalation when limits are reached.
Engram: temporal knowledge graph with fact extraction, conflict detection, and consolidation. Long-running agents maintain accurate context. Works via MCP with any framework.
Your agents crash. Ours recover.
Every step is checkpointed as it happens. When a worker dies mid-run, the scheduler reclaims the lease and resumes exactly where it left off.
$ jamjet run research-pipeline.yaml ▸ Starting execution exec_7f3a... ▸ [Plan] ✓ completed 420ms ▸ [Research] ✓ completed 1.2s ▸ [Analyze] ✗ worker crashed ▸ Lease expired · reclaiming... ▸ [Analyze] ✓ resumed 890ms ▸ [Review] ✓ completed 650ms ▸ [Synthesize] ✓ completed 1.1s ▸ Execution complete · 5/5 nodes · 0 events lost
The Analyze step fails unexpectedly.
The scheduler detects the failure and reclaims work.
No rerun of completed steps. Picks up exactly where it stopped.
All 5 nodes complete. Full execution integrity preserved.
Production outcomes,
not feature lists
Recover from crashes without rerunning everything
Completed steps are never repeated. JamJet resumes from the exact point of failure instead of forcing a full rerun.
Pause safely for human approval
Approval gates suspend durably and survive restarts. Long-running workflows stay reliable even when humans are in the loop.
Replay any execution to debug exactly what happened
Replay a run from checkpoints, inspect traces, and see per-node cost and latency. No guessing from logs.
Connect tools and external agents with MCP + A2A
Use built-in client and server support for both protocols. Connect to tools, delegate to agents, and keep every call inside a durable execution model.
Enforce cost and iteration limits in the runtime
Cost caps, token budgets, and iteration limits are enforced by the runtime, not left to application discipline.
Evaluate agents like software, not vibes
Run evals as workflow nodes with judge-based, assertion-based, latency, and cost scoring. Fail CI on regressions.
Route to the best agent automatically
The Coordinator discovers agents at runtime, scores them on capability, cost, and latency, and routes to the best fit. Optional LLM tiebreaker when scores are close. Every routing decision lands in the event log.
Use agents as tools — three modes
Wrap any agent as a callable tool. Sync for quick tasks. Streaming with early termination for long-running work. Conversational for multi-turn refinement — all with budget enforcement.
Deploy with enterprise controls when needed
Tenant isolation, PII redaction, OAuth delegation, mTLS federation, and retention controls — enforced at the runtime layer.
Your language.
Production-grade
from day one.
Agent strategies, workflow orchestration, cost guardrails, crash recovery. Define agents and workflows in Python or Java — the Rust runtime handles the rest.
from jamjet import task, workflow, approval @task(model="claude-sonnet-4-6", max_cost=0.50) async def analyze(data: dict) -> Report: """Analyze data — checkpointed, cost-capped.""" @workflow async def pipeline(raw: dict): report = await analyze(raw) # crash-safe await approval(report) # durable pause return await publish(report) # resumes here
One question decides your authoring approach.
All three compile to the same IR and run on the same Rust runtime. Pick the one that fits how your agents think.
Agent + Workflow SDK
Multi-agent systems with specialist agents. Each gets a strategy: react, plan-and-execute, critic. Typed Pydantic state. Human-in-the-loop.
from jamjet import Agent, Workflow, tool YAML Workflows
Tool pipelines, MCP orchestration, routing. Change flows without code deploys.
jamjet run workflow.yaml Python Decorators
Custom logic inside nodes. Full Python library access. Class-based workflows.
from jamjet import workflow, node All three produce the same IR → same Rust runtime → same durability, replay, and audit trails. Full guide →
Coming from LangGraph?
Typed state, workflow steps, routing, and graph-like structure — without relearning how to think.
No optional checkpoint plumbing. Every step is durable when you run on the JamJet runtime.
Typed schemas catch state and interface problems before they spread.
Budgets and iteration caps live in the scheduler, not in scattered guard code.
At a glance
Also migrating from:
CrewAI → OpenAI Agents SDK →Built for your role
Ship agents that survive production
Use your Python mental model, but get durable execution, replayable traces, and runtime-enforced safety. Start with @task, scale to full workflows without rewrites.
Govern agents before they reach production
Policy enforcement, audit trails, cost controls, tenant isolation, PII redaction, and OAuth delegation — enforced at the runtime layer, not scattered across application code.
Enterprise controls →Run reproducible experiments
The same runtime properties that make agents reliable in production make experiments reproducible in research. ExperimentGrid, checkpoint replay, publication-ready export.
Explore the research toolkit →Reliable agents for production.
Reproducible agents for research.
Same runtime.
Sweep 6 strategies across models and seeds in one command. Parallel execution with durable checkpoints across every condition.
Replay the exact failed run. Fork with modified inputs for ablations. No need to rebuild infrastructure or re-run completed conditions.
Paper-ready LaTeX tables with mean ± std, CSV, and JSON. From experiment to results without custom scripts or manual formatting.
Six orchestration patterns.
One runtime.
Every major multi-agent pattern is a first-class primitive — not a workaround. Pick the pattern that fits your problem. JamJet handles the durability, observability, and governance.
Single Agent
@task decorator. 3 lines of Python. Best for simple prototypes and single-purpose tasks.
Sequential Pipeline
Chain nodes in a WorkflowGraph. Each step depends on the previous. Best for ETL, document processing, multi-stage analysis.
Parallel Fan-Out
ParallelNode runs branches concurrently. Real async — no GIL. Best for multi-source research, batch classification.
Loop & Critic
EvalNode scores output, retries with feedback if quality is low. Best for code review, content generation, anything quality-critical.
Coordinator
CoordinatorNode discovers agents, scores on capability/cost/latency/reasoning modes, and routes dynamically. Async LLM tiebreaker when scores are close. Full scoring in the event log.
Agent-as-Tool
agent_tool() wraps any agent as a callable tool. Sync for quick tasks, real-time incremental streaming with budget guard and idle timeout, conversational for multi-turn refinement.
Java first, not Java later.
On Maven Central today. Agent runtime, memory layer, Spring Boot starter, and LangChain4j integration —
all under the dev.jamjet group. Drop-in dependencies, zero custom repositories.
<dependency>
<groupId>dev.jamjet</groupId>
<artifactId>jamjet-runtime-spring-boot-starter</artifactId>
<version>0.1.1</version>
</dependency>Kill your process. Restart. Resume from last checkpoint. No sidecar, no Docker. Read the launch post →
In-process method calls instead of REST hops. Virtual threads for 1M concurrent agents. Zero connection pool contention. Benchmark details →
Spring AI, LangChain4j, Google ADK — add @DurableAgent to your existing agent, keep everything else.
Fact extraction, hybrid retrieval, MCP client+server, plugin hot-reload. Full agent runtime in one JVM.
What's new
Pure-Java agent runtime with @DurableAgent crash recovery, ByteBuddy instrumentation,
virtual threads, MCP client+server, and plugin hot-reload. 8.9x faster than REST sidecar.
6 modules on Maven Central under dev.jamjet:jamjet-runtime-*.
engram-spring-boot-starter:0.1.1 ships auto-configuration, property
binding via engram.*, and an optional Actuator health indicator.
Add one dependency, inject EngramClient, done.
Fact extraction, conflict detection, hybrid retrieval (vector + FTS5 keyword + graph),
token-budgeted context assembly, and a 5-operation consolidation engine.
Published as jamjet-engram on crates.io and included in
dev.jamjet:jamjet-sdk on Maven Central.
A landscape survey of every option Java developers have for adding persistent memory to AI agents — LangChain4j, Spring AI, Koog, Embabel, Google ADK, and the rest. Why every option stops at chat history.
Read the landscape →One dependency gives your Spring AI app crash recovery, audit trails, replay testing, and human-in-the-loop gates.
Read more →Need memory first?
Start with Engram.
Open-source memory for MCP-native AI agents. Use it standalone today with Claude, Cursor, or any MCP client — then add JamJet governance when your agents move toward production.
"mcpServers": {
"memory": {
"command": "engram",
"args": ["serve", "--db", "memory.db"]
}
}11 MCP tools · Ollama / OpenAI / Anthropic / Google · SQLite or Postgres
Built for regulated industries
All enforced at the Rust runtime layer, not by convention. Open source, cloud-agnostic, no lock-in.
Security & enterprise docs →Make your agents production-safe.
Start with a 60-second quickstart. Add policy controls, audit trails, and governance as you need them.