Agent Identity Broker

What is it

An Agent Identity Broker is middleware that sits between AI agents and the services they call, issuing scoped, short-lived, auditable credentials instead of letting agents borrow a human's full-access API key. Technically, it's an OAuth 2.1 / token-exchange layer (RFC 8693) plus policy engine: when an agent wants to read a customer record or move money, the broker checks the delegating human's consent, the agent's declared task, and the resource's policy, then mints a token with the minimum scope and a TTL measured in minutes.

The business significance is bigger than plumbing. Today most agent deployments work by handing an LLM a long-lived API key — the classic "confused deputy" problem, where the agent can do anything the key can do, regardless of what the user actually asked for. That's fine in a demo and unacceptable in a bank. The broker is the control plane that makes agentic AI auditable: who delegated what, to which agent, for how long. Whoever owns that layer owns the trust relationship between enterprises and every agent they deploy — a strategically valuable seat, similar to where Okta sat during the SaaS boom.

Why now

Three forces converged in late 2025 and early 2026. First, agent frameworks matured past toy stage: LangGraph, CrewAI, OpenAI's Agents SDK and Anthropic's tool-use APIs made multi-step autonomous agents routine, which means routine over-privilege. Second, the security community named the problem loudly — Simon Willison's "lethal trifecta" framing (private data + untrusted content + external communication) went mainstream, and OWASP's LLM Top 10 now lists excessive agency as a top risk. Third, regulators moved: the EU AI Act's transparency obligations phase in through 2026, and US federal agencies began asking vendors how agent actions are attributable.

The timing is precise because the pain only appears at scale. A single developer scripting one agent doesn't need a broker. A company running 200 agents across 40 SaaS tools, each holding a static key, has an unbounded blast radius — and that company is now the median enterprise buyer. The window is open because incumbents (Okta, Auth0, Microsoft Entra) have identity primitives but no agent-native delegation model yet, and startups haven't filled the gap. Two years from now this will be a feature in every IdP; today it's a product.

Market Evidence

The signal is thin but directionally clean: 3 independent mentions across Google News, Dev Community and Show HN, first surfacing 2026-11-09, with a 100% growth rate off a tiny base. Trend score of 72/100 against 3 total mentions tells the real story — the topic is hot (agent security is discussed constantly), but the specific product category is barely named yet. That's exactly the profile of an early, pre-consensus opportunity rather than a fad.

Cross-platform corroboration matters more than raw volume. HN threads about agent security routinely hit hundreds of comments; Dev Community posts on OAuth for LLM tools are multiplying; and news coverage of agentic AI incidents (prompt-injection breaches, runaway agents) is now a monthly occurrence. The mentions clustering around a named category — identity brokering — suggests practitioners are starting to describe the same missing primitive independently.

My read: this is real demand with weak search signal. Developers feel the pain (they're wiring up API keys by hand every week) but don't yet Google "agent identity broker" because the category name hasn't standardized. That's an opportunity for whoever defines the vocabulary — and a warning that you can't rely on inbound SEO to find customers for the first year.

Who's Behind It

The whales are the identity incumbents. Okta (with Auth0) has the OAuth infrastructure and enterprise trust but is optimized for human SSO; Microsoft Entra ID is bundling agent identity into Copilot's ecosystem, which gives it distribution but locks it to Azure. Google's Agent Development Kit and its emerging "agent identity" docs signal it wants to own the standard. On the security side, Palo Alto, CrowdStrike and Wiz are all shipping AI-security modules that will eventually touch this.

Below the whales sit the standards bodies and loud individuals: the IETF's OAuth working group (token exchange, rich authorization requests), the OpenID Foundation's work on agent identity, and security researchers like Willison whose writing shapes developer consensus. The startup layer is nearly empty — a handful of stealth companies and open-source experiments, nothing with traction.

Competitive dynamic to watch: whoever wins the spec wins the market. If Microsoft and Google converge on a shared delegation profile, a proprietary broker becomes a complement, not a platform. Your defensibility has to come from being the best implementation and policy UX, not from inventing a protocol.

TAM & Market Size

Buyers split into three tiers. Tier one: mid-market and enterprise SaaS companies deploying customer-facing agents — think 5,000–50,000 globally, each with a security budget line already funding Okta, Vanta, and SIEM tools. Tier two: AI-native startups building agents for regulated verticals (fintech, health, legal) that need auditability to close enterprise deals — tens of thousands of them. Tier three: platform vendors (agent builders, RPA successors) who'd white-label a broker via API.

Price tolerance is set by comparison to identity and security spend, not developer tools. Companies pay Okta $2–15 per user per month and Vanta $10k–40k annually without blinking. An agent-identity control plane that prevents a breach is easy to justify at $500–$3,000/month for mid-market and $50k+/year for enterprise. Agent counts, not human counts, become the billing axis — a company with 500 agents at $2/agent/month is a $12k/year account.

Honest caveat: the demand score of 0/100 reflects that measured search and purchase intent for this exact category is effectively zero today. The TAM is real but latent. You are selling a solution to a problem buyers recognize before they recognize your category name — which means outbound and content education, not paid acquisition.

Competitive Landscape

Direct competition is sparse. A few open-source projects implement OAuth token exchange for LLM tools; some agent frameworks (LangChain, CrewAI) ship basic credential management. None offer a standalone broker with policy, audit and consent UX. The real threat is adjacent incumbents bolting this on: Okta's "AI Agents" roadmap, Microsoft's Entra Agent ID, and Auth0's foray into machine identity. They have distribution and trust; they lack speed and agent-native design.

Your differentiation has three lanes. First, developer experience: a 10-line SDK that replaces hardcoded keys is worth more than a feature-complete enterprise console at launch. Second, policy granularity: per-task, per-resource, per-time-window scopes that generic IdPs won't ship for years. Third, auditability as a product: a human-readable "what did this agent do and on whose behalf" log that compliance teams can hand to auditors.

Time horizon: assume 12–18 months before a major IdP ships a credible agent delegation feature. That's enough to win a few hundred design-partner customers and establish the category vocabulary — not enough to coast. Ship fast, publish the spec openly, and make integration with Okta/Entra a feature, not a threat.

Business Model

Recommendation: usage-based SaaS with a free developer tier, sold bottom-up. Free: up to 3 agents, 10k token issuances/month — enough for a solo dev to wire it in and feel the value. Team: $299/month for 25 agents and audit retention; Growth: $999/month for 100 agents plus SSO and policy templates; Enterprise: $50k+/year with on-prem policy engine and compliance reporting. Bill on active agents and token issuances, because that maps to the customer's own cost model and grows with their deployment.

Why this fits: the buyer is a security-conscious engineering lead who wants to try before a procurement call, then needs a number that survives a CFO review. Usage pricing does both. A per-seat model fails because agents outnumber humans 10:1.

12-month forecast: conservative $8k MRR (30 paying teams, no enterprise); base $35k MRR (100 teams, 3 enterprise pilots converting); optimistic $90k MRR (250 teams, 8 enterprise contracts, one platform partnership). CAC estimate: $400–$900 via content and community for self-serve; $8k–15k for enterprise (sales-assisted). Payback: 2–4 months self-serve, 6–9 months enterprise. The self-serve motion is where the model lives or dies — keep onboarding under 15 minutes.

MVP Blueprint

Scope to 5–7 days, one engineer. Core features only:

1. Delegation flow: human OAuth-consents once, broker issues a signed delegation grant tied to that user and a declared task.
2. Scoped token minting: agent requests a token with resource, action, ttl; broker validates against policy and returns a short-lived JWT.
3. Policy file: YAML rules — agent: support-bot, resource: crm.contacts, action: read, max_ttl: 15m, requires_consent: true.
4. Audit log: append-only record of every issuance — who delegated, which agent, what scope, when, outcome.
5. SDK: Python and TypeScript clients with a 10-line integration replacing a static API key.

Cut: admin UI (ship a CLI and a read-only log page), multi-tenant billing (Stripe checkout link), SSO (email magic links), any agent framework integrations beyond one example.

Stack: TypeScript on Node/Fastify or Go; Postgres for grants and audit; Redis for token revocation lists; deploy on Fly.io or Railway. Use an existing JWT library and standard OAuth libraries — do not hand-roll crypto. Ship as API + SDK first, hosted SaaS second. Launch on Show HN with a working demo that shows an agent being blocked from an out-of-scope action — that's the moment that sells it.

Commercial Opportunities

1. Compliance-ready audit trail for regulated AI. Sell a broker plus a "compliance export" that maps every agent action to EU AI Act Article 12 logging requirements and SOC 2 evidence. Target: fintech and healthtech startups closing enterprise deals. $1,500–$5,000/month. Beats generic observability tools because it speaks the auditor's language.

2. White-label broker API for agent platforms. Agent-builder vendors (vertical copilots, RPA successors) need delegation but won't build it. Offer an embeddable API with revenue share. $3k–$20k/month per platform. Beats building your own end-user product because one integration reaches hundreds of their customers.

3. Managed agent security review service. A productized audit: review a company's agent credentials, migrate them to scoped tokens, deliver a report. $10k–$25k per engagement, 2–3 weeks. Beats pure SaaS early because it funds development and teaches you exactly what buyers fear — then converts to subscription.

Product Ideas

🥇 BrokerKit — "OAuth for agents, in 10 lines." A hosted identity broker with SDK, policy YAML, and audit log aimed at AI-native startups. Why now: they're shipping agents this quarter and hitting the static-key wall today; the category name doesn't exist yet, so a clean DX wins.

🥈 AgentLedger — "Every agent action, attributable." A compliance-first audit and consent layer for enterprises deploying customer-facing agents, with EU AI Act and SOC 2 mapping. Target: security and compliance leads at 200–2,000-person companies. Why now: regulatory deadlines create budget; incumbents' logs don't distinguish delegated agent actions.

🥉 Delegation Gateway for Platforms — a white-label API letting agent-builder platforms offer scoped credentials to their own customers. Target: vertical copilot vendors and RPA successors. Why now: they need it to pass their customers' security reviews, and building it in-house distracts from their core product.

SEO Opportunity

Search volume is nascent and will grow steeply as the category names itself — expect "agent identity" and "AI agent OAuth" queries to compound through 2026–2027. SEO difficulty is effectively zero today: almost no page targets the exact intent.

Long-tail targets: "OAuth for AI agents," "AI agent API key security," "confused deputy LLM," "scoped tokens for agents," "agent delegation audit log." Content strategy: publish the reference implementation and a spec explainer, not marketing fluff — developers link to working code. Own the vocabulary before the incumbents do.

Risk Assessment

Risk 1 — Platform absorption. Okta, Microsoft or Google ships agent delegation as a bundled feature, and buyers default to "we already have Entra." Mitigation: be the integration layer across all IdPs, not a replacement; win on policy UX and audit depth.

Risk 2 — Category never crystallizes. Developers keep hand-rolling key rotation and accept the risk; no budget line forms. This is the biggest threat given demand score of 0/100. Mitigation: sell the outcome (passing a security review, closing an enterprise deal), not the category.

Risk 3 — Execution drag. Enterprise sales cycles stretch, compliance features balloon, and a 7-day MVP becomes a 7-month platform with no revenue.

Validate cheaply: publish the spec and SDK on GitHub, run 15 customer-discovery calls with security leads at agent-deploying companies, and pre-sell 5 design-partner slots at $500/month. If fewer than 3 of 15 calls describe the static-key problem unprompted, walk away. If design partners convert within 60 days, double down.

Action Plan

Today: write a one-page spec for the delegation flow (grant → scoped token → audit event), publish it on GitHub, and post it to Show HN and one security community. Gauge comment quality, not upvotes.

Week 1: build the MVP — delegation flow, policy YAML, audit log, Python SDK. Record a 90-second demo of an agent being blocked from an out-of-scope action. Line up 15 discovery calls via LinkedIn and security Slack communities.

Month 1: convert 5 design partners at $500/month (discounted, contract-for-feedback). Ship the TypeScript SDK. Publish two technical posts: "Why your agent shouldn't hold your API key" and "Implementing RFC 8693 for LLM agents." Target: 20 GitHub stars, 5 paying accounts.

Month 3: launch self-serve pricing ($299/$999 tiers), integrate with one major IdP (Okta first), and start the compliance-export feature. Goal: $8k–$12k MRR, 30 accounts, one enterprise pilot in motion. If MRR is under $3k and design partners aren't renewing, reassess before building more.

Related Terms

Confused deputy problem — the security flaw at the heart of this category: an agent acting with more privilege than its delegator intended. The broker exists to fix it.

OAuth token exchange (RFC 8693) — the standards primitive brokers build on; mastering it is table stakes.

Agentic AI governance — the compliance and policy layer (EU AI Act, NIST AI RMF) that turns agent identity from a dev concern into a budgeted enterprise requirement.