Permit

Permit0 is betting on a real gap: identity layers like Okta control which agents exist, observability tools report what already happened, and platform vendors like Salesforce Agentforce or Microsoft Agent 365 only govern what runs inside their boundary. None of them decide whether a specific cross-platform action gets to execute, in real time, before it does. That's the slot Permit0 occupies, and its architecture reflects it. The strongest part is the determinism story. The decision path is rules and cryptography with no model in the hot path — the optional LLM reviewer in the two-stage flow can only deny or escalate, never approve. That asymmetry matters for regulated buyers: you get the flexibility of an LLM second opinion without handing it the ability to wave an $8,500 refund through. Capability tokens in PASETO v4.public format bind approved actions to argument hashes, so a prompt injection can't reuse a signed token against a different payload. Every block, allow, and escalation produces a signed DecisionRecord that's replayable and exportable — which is exactly what a SOC 2 or SR 11-7 reviewer wants to see. The taxonomy is the quieter strength. Twenty-one action domains with opinionated risk defaults means day one isn't an empty policy file — the vendor frames this explicitly as "every policy engine ships an empty policy, and that's the layer Permit0 ends." For a platform team running LangChain, CrewAI, and AutoGen side by side, the Tool Action Compiler normalizing all of them into one canonical ActionSpec is the difference between one policy and three. The weaknesses are honest ones. This is authorization, not observability — it won't tell you what an agent did yesterday at 3am unless that decision went through the gate. It's not an agent builder, and it won't help a solo developer prototyping a chatbot that never executes a tool call. The product is in beta with a design partner program aimed at fintech, insurtech, and healthtech, which means early adopters get founder access but also absorb early-stage risk on roadmap and support. The architecture spec itself is still a draft in progress, with the full document available only via a 30-minute call with the founders. Where it fits: production agents that touch money, PII, or regulated workflows, run across multiple frameworks, and need pre-execution evidence. Where it doesn't: simple chatbots, post-hoc analytics, and teams outside regulated verticals who won't use the compliance mappings that justify the paid tiers.