Open-source primitives for AI-first applications.

We build the missing infrastructure layers between web applications and AI agents. Every project here is agentic-first — designed for a world where AI is a first-class consumer of software, not an afterthought bolted on top.

The harness, not the model. The prompt is the job ticket; The Machine is the operating system — durable state, a dumb deterministic driver, fresh workers, and verification against reality instead of self-report.

Everything else in the stack is a deployment of this pattern, or verifies against it. machine-driver and Conductor are its two reference drivers; Maintainer Gate Blueprint and Agent Control Plane build governance and proof on top of it.

python -m kit score <path-to-deployment-repo>   # dated, evidence-cited L0–L5 packet
python -m kit matrix                            # render the test matrix

Conformance is run, not asserted. A deployment earns a level — L0 Look-Alike through L5 Trusted Autonomy — via the kit; it doesn't claim one.

Read the spec

Producer-side rendering for AI agents. Like i18n, but the target locale is "agent."

Web apps already know their intent, their data, their permissions, and their affordances — then they throw it all away to render pixels. AVL flips this: every page ships a parallel agent-native view alongside the human HTML. Same data, same auth, different rendering target.

/dashboard        → human view (HTML)
/dashboard.agent  → agent view (text/agent-view)

npm install @frontier-infra/avl

MCP is the hands. AVL is the eyes.

Read the thesis · Read the spec · Get started

The Machine is the govern tier everything else plugs into. The rest of Frontier Infra covers declare, behave, enforce, operate, and prove — a full loop for reliable, long-running agent work:

• AVL — Agent View Layer (declare) — a parallel agent-native view shipped alongside the human HTML for every page.
• ADL — Agent Discipline Layer (behave) — layered guardrails plus a goal contract that Warden verifies against reality before a Claude Code session can end.
• Proctor (enforce) — the machine-level engine behind ADL: makes an agent's "done" a verified fact instead of a claim.
• machine-driver and Conductor (operate) — the two reference drivers for The Machine: one drives code work, the other drives multi-agent ops triage.
• Maintainer Gate Blueprint (govern) — reusable ops model for repos where AI agents work in parallel: patrol loops, PR quality gates, machine-checkable handoffs.
• Agent Control Plane / AAR (prove) — Agent Attestation Record: portable, signed, ground-truthed proof of what an agent actually did.
• Roundtable (decide) — convene a council of frontier models — Grok, Codex/OpenAI, GLM, MiniMax, Claude, Gemini — from the CLI or any MCP harness.

See it all running together in stack-demo — a live board whose every claim links a cryptographic receipt you can verify yourself.

• The agent is not a scraper. Applications should render for AI agents the same way they render for non-English speakers — with a parallel, first-class view owned by the producer, not reverse-engineered by the consumer.
• Same session, different render. AI agents inherit existing human sessions. No new auth surface, no new API keys, no second permission model to maintain.
• Ship small, compose well. Every package here is zero-dependency where possible, framework-agnostic at the core, with thin adapters for specific runtimes.
• Start cheap, get value immediately. Conformance ramps (L0 → L3) let teams ship value in hours and expand incrementally.
• Witness, don't trust self-report. Evidence comes from tool output and signed, checkable records — never from an agent's own account of itself.

frontier-infra is the open-source arm of Jason Brashear — 30+ years in infrastructure (Cisco CCIE, Red Hat CE, Dell, Apple, AMD), now building at the intersection of AI agents and web infrastructure.

Writing about the agentic-first future at Frontier Operations on Substack.

• Argent OS — AI agent operating system. AVL is the substrate Argent OS uses to navigate any host application.
• AInode.dev — AI-native development platform.

We're looking for early adopters, framework implementations, and real-world feedback.

• Use AVL in your Next.js app and tell us what breaks
• Build an adapter for SvelteKit, Remix, Nuxt, or Rails
• Write an agent that consumes .agent URLs natively
• Try the discipline layer — install ADL and Proctor in a Claude Code repo and see what they catch
• Join the conversation in Discussions

Infrastructure for the agent-native web.