JamJet: an open safety layer for AI agents (policy, audit, recovery)

JamJet positions itself not as another agent framework but as a runtime safety layer that sits between your agents and models, tools, memory, and other agents—with policy enforcement, checkpoints, approvals, cost limits, and audit-friendly evidence.

Enterprise agent failures rarely come down to “the model wasn’t smart enough.” They come from dull, expensive realities: workers crash mid-run, tools fire when they shouldn’t, spend spirals, or compliance asks what happened and nobody can reconstruct it cleanly.

JamJet describes an open-source (Apache 2.0) approach aimed at exactly that gap: controllable, recordable, and recoverable runs without locking you into a single cloud or framework. The pitch is framework-neutral—keep LangGraph, CrewAI, ADK, OpenAI Agents, Spring AI, etc., and add a layer that can block, wait, record, and resume at the boundary where agents touch the outside world.

What it emphasizes

• Policy hierarchy to stop unsafe tool use before execution.
• Human-in-the-loop gates that survive restarts, with decisions landing in audit trails.
• Crash recovery via checkpointed execution and log replay so work can resume without “just run it again” duplication risk.
• Cost guardrails to halt runaway reflection or spendy loops.
• Audit-oriented evidence (signed, exportable packages in their narrative).
• Memory via Engram (MCP-oriented memory; positioned as usable standalone or with JamJet).
• Polyglot story: Python and Java examples, compiling to a shared IR and running on a Rust runtime (with Java-native runtime claims for embedded JVM use).

JamJet Cloud is presented as a hosted control plane: traces, policy violations, approval queues, cost controls, audit export, and hosted Engram—optional if you want a team “cockpit” rather than only self-hosted pieces.

Why readers should care

If you’re shipping agents past demos, “observability” alone often isn’t enough—you need gates and durable semantics when money, data, or regulation is involved. JamJet frames that as runtime infrastructure rather than prompt tricks.

Caveats for a fair write-up

This is product/marketing positioning plus technical claims; evaluate against your threat model (policy expressiveness, storage of audit logs, tenancy, performance, integrations). Treat benchmarks and multiples (e.g. Java vs sidecar speed) as vendor-reported until you reproduce them.

Further reading

• Site: https://jamjet.dev/