Idea Lab

Pick a vertical. The radar reads everything it tracks there - momentum by year, crowded subniches, dead companies, fresh rounds, grant money, open YC RFS themes - and proposes concrete startups someone should build. The companies below are fictional; the data behind them is not.

Vertical AI agents●Horizontal AI assistants Agent infrastructure●Developer tools AI infra and compute Data for AI Healthcare and bio Fintech●Defense and space Robotics and physical world●Consumer B2B SaaS Commerce and marketplaces Security and compliance●Climate and energy

Agent infrastructure

220

companies, last 2 years

459

tracked all-time

95%

of checked sites alive

Most crowded right now: open source (7) · coding agents (5)

concept 1

Ratchet

Drift firewall that keeps a vendor's API stable for the agents calling it.

Ratchet is installed by the company that publishes an API, not by the agent builder. It watches schema, auth and semantic changes in CI, generates a versioned agent-facing surface (MCP server plus typed tool descriptions) and auto-writes shim adapters plus regression suites so existing agent callers keep working across releases. Product owners get a dashboard of which agent clients would break on the next deploy.

why now: 'Self-Maintaining APIs' is the one YC Fall 2026 RFS theme mapped to this cluster, and the 2026 cohort already shows 106 new agent-infra companies, nearly all of them consuming third-party APIs; supply-side tooling is missing while demand-side integration layers (Executor, Rindler) pile up.
wedge: Series-B vertical SaaS vendors with a public API and a growing share of non-human traffic: start with a free 'agent break report' run against their last 12 releases, then sell the shim generator to the platform team that owns the changelog.
closest real companies: Executor (open source integration management layer) and Rindler (translation layer between agents and the web) both sit on the agent side and patch breakage after the fact; Trieve's death in 2024 shows all-in-one consumer-side API wrappers are hard. Ratchet sells to the API owner and is priced against support-ticket and deprecation cost.
main risk: API vendors may treat agent compatibility as a documentation problem and never fund a paid tool for it.

concept 2

Quorumline

Consensus gate that requires independent model agreement before an agent commits an irreversible action.

Quorumline wraps the write path of an agent system: before a payment, a production config change or a customer-facing email goes out, the proposed action is re-derived by N heterogeneous models and validators, and only committed if a configurable quorum agrees, with dissent logged. It ships as a proxy plus policy DSL, and produces a signed record of who voted what for auditors.

why now: NSF is actively funding the primitive: three 'Resource-Efficient Byzantine Protocols' awards at roughly $403k, $300k and $286k, plus 'Self-Organizing Decentralized Agents' awards near $150k-$200k, while the tracked cluster's average AI-visibility score of 48/100 suggests most agent runtimes still ship with a single-model decision path.
wedge: Mid-size fintech and insurance ops teams that already have a human maker-checker rule they are being asked to automate: replace the second human on one action type (refunds above a threshold) and charge per gated action.
closest real companies: TrustAI (continuous compliance and governance for agents on sensitive systems) audits and reports; Archal and the dead Atla (2023) tune quality offline. Quorumline is an inline commit gate, not observability, and its output is a control artifact an internal auditor can sign off.
main risk: Latency and 3-5x inference cost per gated action make buyers prefer a cheap deterministic rule plus a human on call.

concept 3

Purser

Spend mandates and machine-readable receipts for agents that buy things.

Purser issues scoped payment credentials to individual agent runs: per-merchant, per-amount, per-time-window mandates, with a policy engine that pre-authorizes categories and a receipt normalizer that returns structured line items back into the agent's context. Finance teams see spend by agent, by task and by prompt origin, and can revoke a single agent's authority without touching the company card.

why now: Agent payments sits inside the cluster definition but of the 36 most recent tracked companies only Financial Datasets touches money at all, and it sells market data rather than settlement; 220 companies launched in the last two years are increasingly running purchase and procurement tasks with no per-agent spend primitive.
wedge: SMB and agency ops teams whose agents already buy ads, data credits and SaaS seats: start as a virtual-card issuing wrapper with an MCP tool for 'request budget', sold to the person who currently reconciles those charges by hand.
closest real companies: OneCLI (identity gateway for AI agents) covers authentication and access, not authorization to spend; Decawork's agent control plane targets IT rather than finance; Magma monetizes traces, which is the reverse flow. Purser's differentiator is the mandate-to-receipt loop, not identity.
main risk: Card issuing and money movement bring sponsor-bank and compliance overhead that a three-person team cannot carry before the first real volume.

concept 4

Backout

Compensating-transaction layer that gives enterprise agents a real undo across SaaS systems.

Backout records every write an agent makes into Salesforce, NetSuite, Jira, Workday or a database as a reversible unit with a captured pre-state and a generated inverse operation, then exposes one-click and API rollback at the task, session or fleet level. When a clean inverse is impossible it flags the action as irreversible in advance so the orchestrator can route it to a human.

why now: The 2026 and 2025 cohorts (106 and 99 companies) are dominated by runtimes and control planes - herdr, SpaceFlow, Decawork, Akon Labs - all of which assume a rollback story they do not own, and 95% of tracked sites are still live so the population of agents writing to production systems is growing rather than churning out.
wedge: Enterprise IT teams piloting agents on one system of record: instrument Salesforce writes only, prove a two-minute recovery from a bad bulk update, and sell as a prerequisite for widening the pilot.
closest real companies: Amulet's high-performance file system for agents handles state at the storage layer, Codag compresses logs, Buildbox and Agnost AI observe behavior after the fact. Backout is the only concept here that stores semantic inverses for third-party SaaS APIs; the dead Propolis (2025) shows post-hoc browser QA was not enough on its own.
main risk: Many enterprise APIs cannot be cleanly inverted, so coverage stays partial and buyers keep their existing change-freeze process instead.

concept 5

Cellwright

Network digital twin that tests agent-issued RAN config changes before they touch live cells.

Cellwright gives carrier automation teams a simulated radio access network where an AI agent's proposed parameter changes are replayed against recorded traffic and interference conditions, scored for KPI regression, and either promoted or blocked. It ships as a container with vendor-neutral config adapters and an evaluation harness that produces evidence packs for change-approval boards.

why now: NTIA's Public Wireless Supply Chain Innovation Fund has an open solicitation specifically for 'Solutions for AI-Native RAN' closing 2026-09-09, and the cluster shows only Simantic (firmware simulation for AI agents) working anywhere near telecom hardware simulation.
wedge: Apply to the NTIA solicitation with one tier-2 or regional carrier as a letter-of-support partner, and start with a single use case: automated antenna tilt and power tuning validated offline before deployment.
closest real companies: Simantic does firmware simulation for AI agents but not network-level KPI regression; Instance does automated evals for robot policies and Coasty does real-world evals for computer-use, both in different physical domains. No tracked company targets RAN operations.
main risk: Carrier sales cycles and vendor-locked OSS stacks stretch past the grant period, leaving the team with a demo and no commercial buyer.

concept 6

Groundwatch

Spatial permission layer deciding where a camera or robot agent is allowed to act.

Groundwatch maintains a machine-readable map of a facility - zones, equipment, occupancy rules - and mediates every action a physical-world agent proposes against that map, denying or escalating anything outside its granted envelope. It logs each decision with the spatial and video evidence used, producing an audit trail for EHS and insurance reviews.

why now: NSF's SmartSpaces award under VINES Track 1 puts roughly $700k into spatial awareness services for new verticals, and the S26 batch alone added Mireye (infrastructure for physical world AI agents), OpenVector (any camera as an autonomous worker) and Waddle Labs (agents for robotics control) - capability suppliers with no shared permission model.
wedge: One warehouse or manufacturing site already running a camera-based agent pilot: start as a read-only 'zone violation' monitor on their existing feeds, then turn on blocking once safety staff trust the alerts.
closest real companies: Mireye and OpenVector build the perception and action stack Groundwatch constrains; Instance evaluates robot policies offline; TrustAI governs agents on sensitive digital systems rather than physical space. The wedge is the site safety owner's sign-off, not the ML team's metrics.
main risk: Industrial buyers already trust PLC interlocks and physical guarding, so a software permission layer reads as redundant paperwork.

Fictional concepts generated 2026-08-17 by claude-opus-5 from MarkosWeb data (accelerator portfolios, funding news, US federal grants, YC RFS). Numbers reflect tracked sources only; treat every concept as a research prompt, not a plan.