Azimuth

Azimuth · context-first methodology

Your agentic workflows are still just automation.

An AI decision you can't explain is a liability. Azimuth makes governance a design-time decision: auditable, named, and defensible to a regulator.

View on GitHub → Read the argument
Knowledge Contract Gate PASS · FAIL · ESCALATE Bound Evidence

The argument

i

The question you can't answer

A loan is denied. A change ships to production. Months later, a regulator asks: show exactly why, and who decided. Most agentic systems can't. The process lives in scattered Python, prompt templates, and a doc that went stale.

ii

It's a design problem

Frameworks execute agent graphs. None of them specify the contract: what each agent must do, where a human must decide (not just approve), and which experts must answer before the workflow runs. You can't bolt that onto a system never designed for it.

iii

Governance is a file you author

Other approaches run at inference time or after the fact: monitoring, guardrails, audit logs. Azimuth runs at design time. One context spec, written once, executes unmodified on any conformant runtime. Governance becomes a file you author, not a system you operate.

The context stack outlives every technology change. It was never about the technology.From the context-first manifesto

Prior art is acknowledged openly. The novelty is naming and packaging the discipline, not claiming the parts are new.

Three primitives

Three primitives, one portable validated spec. No runtime gives you these at design time.

Knowledge Contracts

Expert knowledge, typed

Expert knowledge as a typed artifact (an SME pack): named answerer, role, evidence links, confidence, sign-off. Not a prompt. Inspectable, versionable, auditable.

Gates

Typed contracts

Every workflow transition passes a gate that returns PASS, FAIL, or ESCALATE, owned by a named human role, with an explicit checklist. The workflow cannot advance on an agent's say-so.

Evidence binding

Structural audit trail

Every claim in every artifact cites a source: an upstream artifact section, a data tag, or a specific SME answer. The audit trail is structural, not optional.

How adoption works

You never hand-write YAML. A /skill interviews you and compiles the result.

  1. Step 1

    The interview

    The Consultant skill asks three questions: what decision the workflow produces, where a human must decide with real judgment, and what knowledge lives only in your domain experts.

  2. Step 2

    The agreement

    The answers become a Context Design Document: prose your stakeholders read and sign off. No YAML, no schema knowledge required.

  3. Step 3

    The execution

    It compiles to a validated context-spec.yaml that any conformant runtime runs unmodified.

Status · azimuth/v1alpha

Pre-1.0 reference implementation. Breaking changes are possible until v1. Apache-2.0, patent grant included.

Shipped: this landing page.  Next: the validator toolkit and the Consultant skill.