Build AI agents that
don't lose their work.

JamJet is the durable runtime for AI agents — with checkpoint replay, full execution traces, and runtime-enforced reliability.
Write Python. Run with Rust reliability.

$ pip install jamjet

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.

scheduler Plan Research Analyze Review Synthesize
jamjet · crash recovery
$ 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
1
Worker crashes mid-execution

The Analyze step fails unexpectedly.

2
Lease reclaimed automatically

The scheduler detects the failure and reclaims work.

3
Resumes from checkpoint

No rerun of completed steps. Picks up exactly where it stopped.

4
Zero events lost

All 5 nodes complete. Full execution integrity preserved.

Production outcomes,
not feature lists

Recover from crashes without rerunning everything

Event-sourced execution means completed steps are never repeated. Resume from the exact point of failure.

Pause safely for human approval

Native human-in-the-loop nodes. The process sleeps durably until a human responds — even across restarts.

Replay any execution to debug exactly what happened

Full checkpoint replay with OpenTelemetry traces. Per-node cost attribution. No more guessing.

Connect tools and external agents with MCP + A2A

Both protocols built in — client and server. Connect to any tool, delegate to any agent.

Enforce cost and iteration limits in the runtime

Token budgets, cost caps, iteration limits. Enforced by the Rust runtime, not by convention.

Evaluate agents like software, not vibes

LLM-as-judge, assertion, latency, and cost scorers. Run evals as workflow nodes. CI exit codes on regression.

Deploy with enterprise controls when needed

Tenant isolation, PII redaction, OAuth 2.0 delegation, mTLS federation — enforced at the runtime layer.

Python you know.
Durability you don't
have to build.

Every @task is checkpointed. Every workflow survives crashes. Human approval gates pause durably. Cost limits are enforced by the runtime.

pipeline.py durable workflow
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

Coming from LangGraph?

Same Python mental model

TypedDict → pydantic.BaseModel. StateGraph → Workflow. Familiar patterns, stronger guarantees.

Built-in durability

Not optional checkpoint plumbing. Every step is durable by default, enforced by the Rust runtime.

Per-step validation

Typed schemas at every boundary. Catch errors before they propagate through the graph.

Runtime-enforced limits

Cost caps and iteration budgets enforced by the scheduler, not by your code.

At a glance

Plain Python Easy start, no durability
LangGraph Graph orchestration, optional checkpoints
JamJet Python mental model + durability by default

Also migrating from:

CrewAI → OpenAI Agents SDK →

Built for your role

For Builders

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.

Read the quickstart →
For Platform Teams

Standardize agent infrastructure

One runtime for reliability, observability, eval, and governance. Tenant isolation, PII redaction, OAuth delegation, and mTLS — all enforced at the Rust layer.

Enterprise controls →
For Researchers

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.

Compare

Sweep 6 strategies across models and seeds in one command. Parallel execution with durable checkpoints across every condition.

Replay

Replay the exact failed run. Fork with modified inputs for ablations. No need to rebuild infrastructure or re-run completed conditions.

Export

Paper-ready LaTeX tables with mean ± std, CSV, and JSON. From experiment to results without custom scripts or manual formatting.

Explore the research toolkit

Enterprise-grade when you need it

Tenant isolation PII redaction OAuth 2.0 delegation mTLS federation Retention policies

All enforced at the Rust runtime layer, not by convention.

Security & enterprise docs →

Start building in 60 seconds.

$ jamjet init my-agent
Read the quickstart View on GitHub