Praxis
A Go library for orchestrating LLM agents with the guardrails built into the kernel — typed invocation state machine, policy hooks at every sensitive boundary, budget enforcement and mandatory telemetry. It is the kernel of a three-layer stack.
As of October 2026.
How it works
Context
Most agent libraries are thin wrappers over a provider's HTTP API. That is the right shape for calling a model, and the wrong shape for a team that has to answer for what the agent did, what it cost and who authorised it.
Praxis is the first layer of a three-layer stack. The kernel runs one agent's loop. praxis-forge builds agents from a declarative spec. praxis-os, a control plane for multi-agent systems, is still in design.
Problem
Governance added on top of an orchestration library is governance that can be bypassed. If policy checks, budgets and telemetry live in application code, every new call site is a chance to forget one, and there is no framework-level guarantee to point at during an audit.
Role
Author of the kernel, praxis-forge, the benchmark lab and the control-plane design.
Decisions
Guardrails in the kernel, not in the applicationGuardrails and policy enforcement
A typed invocation state machine with policy hooks and filter chains at every security-sensitive boundary, plus four-dimensional budget enforcement and a typed error taxonomy. The cost is overhead that a simple API call does not need — the README says so plainly and points at the thin SDKs for anyone who wants that instead.
Provider-agnostic LLM interfaceAgent orchestration
Model access sits behind an interface, so the runtime is not bound to one vendor.
Telemetry is mandatoryLLM observability and tracing
OpenTelemetry is not an optional integration but a framework guarantee, so every invocation is traceable without the application opting in. Optional per-call identity signing sits alongside it.
Declarative agents, built reproducibly, in a separate layerAgent orchestration
praxis-forge takes a typed AgentSpec, resolves each component through a typed factory in a registry, validates the composition strictly and builds a BuiltAgent on a Praxis orchestrator. Tools, policies, filters, budgets, credentials, identity and MCP bindings are settled in one deterministic build, and the manifest records a stable hash of the normalized spec. The cost is a second layer to learn. It does no orchestration by design: routing, delegation and long-lived sessions belong to the control plane.
Details
Results
- Apache 2.0, published on pkg.go.dev with CI on the repository.
- The README states what the library is not for, which is the honest way to ship an opinionated runtime.
- praxis-forge is public under Apache 2.0, pre-1.0 at v0.7.0.
- The benchmark lab and the control plane are private, self-reported work.
The kernel enforces budgets in four dimensions — wall-clock time, LLM tokens, tool-call count and estimated cost — and a breach ends the invocation.
you can check this: praxis
OpenTelemetry spans are a framework guarantee, emitted for every invocation without the application opting in.
you can check this: praxis
Released under Apache 2.0 and published on pkg.go.dev, with CI on the repository.
you can check this: praxis
The stack has three layers — the kernel, a declarative build layer (praxis-forge, public) and a control plane in early development (private).
you can check this: praxis, praxis-forge
praxis-forge validates a declarative AgentSpec and builds it into a reproducible BuiltAgent, recording a stable SHA-256 hash of the normalized spec in the build manifest.
you can check this: praxis-forge
A private benchmark lab measures runtime behaviour — tracing, guardrails, concurrency, retries and cancellation — under deterministic local workloads, not model quality. No results are published.
self-reported; no independent source