AgentOS
TypeScript agent framework where agents write their own tools
Overview
AgentOS is an open-source (Apache 2.0) TypeScript framework from Frame (Framers Lab) for autonomous agents that write and sandbox-test their own tools at runtime, coordinate via six multi-agent strategies, and use a neuroscience-inspired persistent memory system across 11+ LLM providers.
Pricing
Pricing shown for reference only. These figures reflect RECATOOLS research as of 3 Sep 2026 and may be out of date or incomplete. This is not financial or purchasing advice — always confirm the current price on the provider’s official website before making any decision.
What you can produce with AgentOS
- Runtime tool generation with LLM-judge approval and node:vm sandboxing
- Persistent cognitive memory (8 neuroscience-backed mechanisms)
- Six multi-agent orchestration strategies
- 11+ LLM provider support with automatic fallback routing
- Multimodal RAG
- Voice/telephony pipeline integration
- Published LongMemEval-S/M benchmark scores
AgentOS bets on two ideas most agent frameworks don't touch: agents that write their own TypeScript tools mid-task (an LLM judge reviews the spec, then it runs sandboxed in node:vm before joining the permanent toolkit) and a memory system built on actual cognitive-science mechanisms — Ebbinghaus decay, retrieval-induced forgetting, reconsolidation — rather than a flat vector store. The project backs this with published numbers: 85.6% on LongMemEval-S and 70.2% on LongMemEval-M, reportedly the highest score among open-source frameworks with reproducible methodology on that harder benchmark. That's an interesting claim worth verifying yourself before betting production infrastructure on it, since it's self-reported rather than independently run. At under 600 GitHub stars and built by a small infrastructure shop (Frame, also behind Paracosm and Wilds.ai) rather than a funded platform company, AgentOS is early and its abstractions are heavier than minimalist options like Mastra. Worth a look for memory-heavy, long-running agents; not yet a safe default choice.
What people say
AgentOS is younger and far less discussed publicly than competitors like VoltAgent or Mastra — there's no G2 or Capterra listing, no visible Reddit or Hacker News threads breaking down real-world usage, and the GitHub repository (framerslab/agentos) sits at 591 stars, a fraction of VoltAgent's 10,000-plus. That doesn't mean the engineering is thin; it means the evidence base is mostly the project's own documentation and benchmark claims rather than independent third-party validation.
The two headline features are unusual enough to be worth explaining on their own terms. Runtime tool generation means that when an agent hits a task it doesn't have a tool for, it writes a TypeScript function with a Zod validation schema, an LLM acts as judge to approve the spec, and the tool then executes inside an isolated node:vm sandbox before it's allowed to join the agent's permanent toolkit for the rest of the session — a novel approach to the "agent needs a tool it doesn't have" problem that most frameworks solve by requiring a human to pre-register every tool.
The memory system is the other differentiator: persistent, cross-session memory implemented via eight named mechanisms lifted from cognitive-science literature, including Ebbinghaus forgetting-curve decay and retrieval-induced forgetting, plus an optional HEXACO personality model meant to shape how an agent weighs decisions without extra prompt engineering.
On benchmarks, the project publishes LongMemEval scores — 85.6% on LongMemEval-S at roughly $0.009 per correct answer using GPT-4o, and 70.2% on LongMemEval-M, which the maintainers describe as the best score among open-source libraries above 65% on that harder variant with reproducible methodology. These are self-reported numbers from the project's own documentation rather than a neutral third-party benchmark run, so they're worth treating as a starting point for your own testing rather than a settled fact.
The framework also covers six multi-agent orchestration strategies, guardrails across five stated security tiers, multimodal RAG, voice/telephony integration and OpenTelemetry-based observability, and it's built and maintained by Frame (Framers Lab, Inc.), a small AI-infrastructure shop that also runs Paracosm and Wilds.ai — not a venture-scale platform company, which matters for anyone weighing long-term maintenance risk.
Summary of public user & expert reviews, compiled by RECATOOLS.
About this listing
This entry was compiled from publicly available data including AgentOS's official website, press releases, documentation, and reputable third-party publications. RECATOOLS is not affiliated with AgentOS unless explicitly stated.
Third-party AI tools update their pricing, features, availability, and policies frequently. Information here may be outdated by the time you read this — we make reasonable efforts to keep listings current, but cannot guarantee absolute accuracy.
For the latest details, please refer to AgentOS directly →
Spotted something out of date? Suggest an update →
Alternatives to AgentOS
More in Agents & Automation
