BETA Design partner program — fintech, insurtech, healthtech teams shipping AI agents into production. Apply →
Preview · Draft in progress

Architecture

The full architecture spec is in active development. Here's the structure of what it will cover when published — and how to get the working draft today.

What this spec will cover

An end-to-end walk through Permit0's engine — how a tool call becomes a canonical action, gets scored, evaluated, and either executed under a signed token or blocked.

01

Engine internals

The full request lifecycle from agent tool call to gated execution. How the SDK intercepts, how the gate communicates with the policy plane, where state lives, and where the hot-path latency budget is spent (target: <5ms p99).

02

Tool Action Compiler

How heterogeneous tool calls — LangChain, MCP, CrewAI, OpenClaw, raw Python — normalize into canonical actions across 21 domains. The 7-stage compilation pipeline, the canonical ActionSpec schema, and how new tools and frameworks plug in without policy rewrites.

03

Risk Engine

The 9-flag × 10-amplifier scoring model. Why risk is intrinsic to the verb (universal baseline) and amplified by context (session, actor, environment, amount, destination). The math behind the score, and how it composes with policy to drive ALLOW / DENY / REVIEW outcomes.

04

Two-stage review

A deterministic scorer makes the first decision. An optional LLM reviewer runs second — and is architecturally permitted only to deny or escalate, never to approve. Why this asymmetry is the only LLM-in-the-loop pattern that doesn't compromise determinism.

05

Capability token design

PASETO v4.public format, payload binding to argument hashes, constraint envelopes, TTL, replay protection, signing infrastructure, and the verification path inside the Execution Gateway. Why an adversarial prompt cannot bypass a signed gate, even with valid OAuth credentials.

06

Policy distribution and DecisionRecord

How policies are authored, versioned, and distributed across deployment modes. The DecisionRecord schema produced for every evaluation — signed, replayable, and exportable. How the audit trail composes into evidence for SOC 2, EU AI Act, NAIC, SR 11-7 reviews.

Need this before it's published?

Two paths — pick whichever fits your timeline.

Path 1 · For high-intent readers

Walk through the working draft with the founders.

The founders are happy to share the WIP architecture document on a 30-minute call and answer security questions in real time. We've done dozens of these — they're substantive, not sales calls.

Talk to founders
Path 2 · I'd rather wait

Notify me when published.

One email when this document ships. No newsletter, no marketing cadence — just the link to the published doc.

We'll only use your email to send this single notification.