Week of Oct 9, 2026 Batch SaaS

Agent identity and runtime authorization

Startup Idea Brief Published Week of Oct 9, 2026

A real-time policy engine that approves or blocks every AI agent's tool call, data scope, and time window, and logs why.

Every enterprise now has agents calling internal APIs, databases, and SaaS tools on their own, often in chains that span dozens of steps and multiple vendors. Identity providers can tell you an agent exists and give it a credential, but none of them decide, in the moment, whether this specific agent session should be allowed to read a customer's financial records at 2am, call a refund API for the fifth time in an hour, or hand off to a third-party agent it has never talked to before. This product is that decision layer: a runtime authorization engine that sits between every agent and every tool call, evaluates the request against a policy graph built from role, task, data sensitivity, and historical behavior, and returns an allow, deny, or step-up-approval verdict in milliseconds.

It is built for security and platform engineering teams at mid-market and enterprise companies running agent fleets in production, not for the identity and access teams issuing the credentials in the first place, that is the identity provider's job. The wedge is the audit trail: every decision, the policy that triggered it, and the data scope touched gets logged in a format compliance and SOC teams can hand to auditors directly, turning a question nobody can currently answer, what did this agent actually do and why was it allowed, into a report generated on demand.

It plugs into whatever identity provider, agent framework, and tool registry a company already has, so it competes on decision quality and auditability, not on replacing the identity stack underneath it.

Why Now

Agents now run continuously and chain tool calls across systems without a human in the loop at each step, so the old IAM model, where a human authenticates once and keeps a role until someone revokes it, has no mechanism for deciding why a specific action is happening, how long an agent session should retain access, or which of the dozens of tools in its reach it should be allowed to touch right now. Static roles and standing permissions were built for sessions that last hours and are supervised; agent sessions last indefinitely and are not. That gap is exactly where incidents happen, an agent with a broad role takes an action nobody reviewed because no human was positioned to catch it before it executed. Security teams are now asking for policy decisions made at the moment of the tool call itself, not just at login.

Demand Signal

Morgan Stanley's cybersecurity research team estimates the enterprise identity security market could roughly double from about 24 billion dollars today to 60 billion dollars over the next two years, driven specifically by demand for tools that govern non-human and agent access. Y Combinator's Winter 2026 batch includes Agentic Fabriq, pitched as Okta for agents, which handles authentication and least-privilege scopes for agent sessions, evidence that investors already see identity governance for agents as a venture-scale category. CrowdStrike has moved further down the same stack: its Continuous Identity for AI Agents, launched at Identiverse in June 2026 and expanded into a full Agentic Identity Provider at Fal.Con in September 2026, brokers short-lived, scoped tokens and ties every agent action back to a human or workload in real time, confirming that runtime, per-action authorization, not just identity issuance, is now expected from top security vendors.

Defensibility

The core asset is a policy graph that encodes, action by action, which tool calls, data scopes, and time windows are acceptable for a given agent, task, and risk context, built up from every decision made across every connected identity provider, tool registry, and agent framework a customer runs. That graph gets more accurate and harder to replicate the longer a customer runs real agent traffic through it, because it is trained on the customer's actual policy exceptions, incident history, and auditor feedback, not a generic ruleset. Once compliance and audit teams build their SOC 2, HIPAA, or FedRAMP evidence packages around this system's logs, ripping it out means redoing that evidence trail, which creates real switching cost on top of the engineering cost of rewiring a decision layer that sits in front of every tool call.

Market Signals

Identity security market to double: Morgan Stanley estimates enterprise identity security spend could grow from roughly 24 billion dollars to 60 billion dollars over two years as governing non-human and agent access becomes its own budget line. Morgan Stanley

YC-backed Okta for agents: Agentic Fabriq, a Y Combinator Winter 2026 company, launched an identity and permissions control plane for AI agents with least-privilege scopes and signed audit logs. Y Combinator

CrowdStrike ships real-time agent authorization: CrowdStrike's Continuous Identity for AI Agents and its follow-on Agentic Identity Provider broker short-lived, scoped tokens and evaluate agent actions in real time, launched in June and September 2026 respectively. CrowdStrike

Non-human identity governance already a budget line: Independent market research from Meticulous Research pegs the non-human identity access management market at roughly 12.2 billion dollars in 2026, growing at about 12 percent a year, confirming NHI governance spend is already tracked as its own category distinct from general identity security. Meticulous Research

How I'd Build This

MVP Scope

Build a policy decision point that sits as a lightweight proxy or SDK between an agent orchestration framework, such as LangGraph, CrewAI, or a custom in-house agent runtime, and its tool-calling layer, evaluating each call against a small set of declarative rules covering which tools, which data scopes, and which time windows a given agent role may use. Ship it with a structured, exportable audit log from day one, since that log is the artifact design partners will actually show their compliance team.

First Customers

Target the security and platform engineering leads at mid-market fintech, healthcare, and insurance companies already running agents in production for support, claims, or back-office workflows, since they face the most immediate audit pressure. Find them through CISO and platform-engineering communities, SOC 2 and compliance consultancies who see this exact gap in client audits, and by showing up at identity-security events like Identiverse and Gartner IAM, where CrowdStrike and Okta's own launches are already priming buyers to look for this category. Ten design partners is realistic within 90 days through direct outreach plus warm intros from compliance consultants.

Moat Sequence

First, ship the real-time decision engine and audit log so design partners run real agent traffic through it. Second, build the policy graph that learns from each customer's actual approvals, denials, and incident history so the ruleset gets sharper than any generic template. Third, package the audit trail into pre-built compliance evidence formats for SOC 2 and HIPAA so switching away means rebuilding an auditor relationship, not just a config file.

90-Day Timeline

Months 1 to 3: ship the MVP decision point integrated with one or two agent frameworks and one identity provider, and land three design partners. Months 3 to 6: harden the policy graph with real customer decision data, add compliance evidence templates, and convert design partners into paying customers.

Building in this space?

I've taken an agentic AI product from first build to acquisition. If you're exploring this idea or one close to it, a 30-minute call is a good place to pressure-test it.

Book a 30-min call

Sources

Checking IDs for AI Agents, Morgan Stanley Agentic Fabriq: Okta for Agents, Y Combinator Launch CrowdStrike Unveils Continuous Identity for AI Agents Non-Human Identity Access Management Market Report, Meticulous Research