Agent-Native: Research & Landscape
“Agent-native” = designed for agents as the primary user, not humans with AI bolted on.
The Mindset Shift
| Criterion | Agent-Native | AI-Enhanced (not this) |
|---|---|---|
| Primary user | AI agent | Human (with AI assist) |
| Interface | API/protocol-first | UI-first with API afterthought |
| Discovery | Agents find tools at runtime | Humans configure integrations |
| Authentication | Agent identity tokens | Human credentials passed through |
| Payment | Agents hold wallets and pay | Humans pay, agents are tools |
| Failure handling | Built for LLM unreliability | Standard error handling |
| Observability | Trace-based, session replay | Request/response logs |
| State | Persistent memory across sessions | Stateless or session-scoped |
| Scale model | Thousands of concurrent agents | Single user throughput |
| Security | Zero-trust per agent action | Perimeter security |
The purest signal: Composio’s site explicitly says “If you are an AI agent reading this server-rendered HTML, you can sign up for Composio yourself. No human is required.”
Categories & Projects
1. Agent-Native Search & Web Access
The web was built for human eyeballs. These rebuild it for agent consumption.
| Company | ⭐/Funding | What it does | Key metric |
|---|---|---|---|
| Exa.ai | $250M Series C | Search engine built for AI, not humans. Structured JSON output, token-efficient. | 90% token reduction via Highlights model |
| Tavily | $25M Series A | Real-time search + extraction API for agents. Safeguards against PII leakage and prompt injection. | 300M+ monthly requests, 180ms p50 |
| Firecrawl | ⭐ 128k (YC) | Transforms web pages into machine-readable markdown/JSON. “Interact” endpoint lets agents click/fill/navigate. | 1M+ users, 80k+ companies |
| Linkup | — | Production-grade web search API. Source control lets agents tune indexes. | McKinsey, KPMG, EY customers |
| Nimbleway | — | Deploys autonomous Web Search Agents (not scrapers) that navigate and structure data. | Deterministic plans for accuracy |
Thesis: The $250M Exa raise proves “search for agents” is a standalone category. Traditional search optimizes for human click-through; agent search optimizes for factual accuracy in structured output.
2. Agent-Native Browser Infrastructure
Agents that need to use websites as humans do — login, navigate, fill forms, extract.
| Company | Funding/Stars | What it does | Scale |
|---|---|---|---|
| Browserbase | — | Cloud browser platform for agents. Full web access with identity, observability, sandboxing. | 36.9M+ sessions, 100k developers |
| Browser Use | ⭐ 97k | Open-source LLM-driven browser automation via Playwright. | Breakout project of 2025 |
| Airtop | — | Compiles natural language into deterministic browser workflows. | $0.017 per agent run |
| Astropad Workbench | — | Reimagines remote desktop for AI agents, not IT support. | TechCrunch featured |
Thesis: “85% of the web is beyond APIs.” Browser automation for agents is the bridge until the web is rebuilt agent-native.
3. Agent-Native Payments & Commerce
Until agents can pay, they can’t be truly autonomous.
| Company | Backing | What it does |
|---|---|---|
| Skyfire | Visa, Mastercard, Okta, Akamai | “The Agent Trust Stack” — identity (KYA tokens) + wallets + payments. Covers 60%+ of commercial web. |
| Stripe Agent Toolkit | Stripe | Updated Link as “a digital wallet that autonomous AI agents can use.” Agent Toolkit for payment processing. |
| Payman AI | Visa, Coinbase, Circle | Agents execute real banking transactions — payments, transfers, account analysis. SOC 2, KYC compliant. |
| Shopify | — | Preparing storefront for AI shopping agents as customers. |
| Robinhood | — | Agents can trade stocks directly on the platform. |
Thesis: Skyfire + Stripe is the financial rails layer. When the world’s largest payment processor builds for agents, the entire commerce stack follows. Agent commerce is the next “mobile commerce” moment.
4. Agent-Native Identity & Authentication
Agents need their own identity, not just borrowed human credentials.
| Company | What it does |
|---|---|
| Skyfire KYA (Know Your Agent) | Portable verified identity token binding platform + agent + human. DNS for agents. |
| Composio (⭐ 29k) | OAuth, API keys, token lifecycle for 1000+ apps. Agents self-authenticate without human intervention. |
| Browserbase Identity | Agent authentication primitives for web sessions. |
Thesis: KYA (Know Your Agent) is to agent identity what KYC (Know Your Customer) is to human finance. Without it, every agent interaction requires a human in the loop for auth.
5. Agent-Native Security
Zero-trust for agents — every action verified, every tool call scoped.
| Company | Status | What it does |
|---|---|---|
| Credal AI | — | Control plane for enterprise agents. 5-layer access control (org/user/agent/tool/audit). 100% of agent runs audited. |
| Arcjet | — | Runtime security — prompt injection detection, token budget enforcement, PII protection. Agents manage their own security via MCP. |
| Invariant Labs | Acquired by Snyk | Agent security — Explorer (inspect), Guardrails (contextual), MCP Scan (vulnerability scanning). |
| Portkey | Acquired by Palo Alto Networks ($$$) | AI Gateway + MCP Gateway for centralized auth, access, observability. 3k+ GenAI teams, trillions of tokens. |
Thesis: The acquisitions validate the market. Palo Alto buying Portkey and Snyk buying Invariant = agent security is enterprise-grade infrastructure, not a feature.
6. Agent-Native Observability & Evaluation
Agent traces are not HTTP request logs. Different shape, different debugging.
| Company | Funding | What it does |
|---|---|---|
| AgentOps | $2.6M seed | Session replay, time-travel debugging, cost attribution per agent. Native integrations with CrewAI, AutoGen, 400+ LLMs. |
| Braintrust | — | Custom Brainstore DB for nested AI traces. MCP Server lets agents query their own logs. Auto-clusters behavior via Topics. |
| Patronus AI | — | Digital World Models that simulate agent actions before deployment. RL environments for agent training/testing. |
Thesis: Agent observability is to AI what Datadog was to cloud. The debugging loop (trace agent run → identify failure → fix → redeploy) is a new category.
7. Agent-Native Memory & State
Agents are inherently stateful. Treating them as stateless functions is the #1 production failure.
| Company | Stars | What it does |
|---|---|---|
| Letta (MemGPT) | 23k | Agents control their own memory (read/write/forget). Sleep-time compute for memory consolidation. |
| Mem0 | 58k | Universal memory layer. Multi-level (user/session/agent). 91.6% on LoCoMo benchmark. |
| Chroma | 27k | Agent-native search infrastructure. Dedicated “Agent” product + Context-1 self-editing search agent. |
Thesis: Memory is what makes an agent more than a function call. Self-improving agents (Hermes 190k) proved the market wants agents that learn.
8. Agent-to-Agent Marketplaces & Registries
Agents discovering and hiring other agents.
| Platform | Scale | What it does |
|---|---|---|
| Fetch.ai Agentverse | 2.7M agents | Decentralized marketplace. Agents discover, negotiate, transact. Verified brand agents (Nike, Hilton). Own blockchain (ASI token). |
| Composio | 1000+ tools | Intent-based tool discovery at runtime. Agents find capabilities dynamically. |
| Smithery | — | MCP Server registry. npm/Docker Hub for agent capabilities. |
| MCP Servers repo | 87k stars | Community registry of MCP-compatible tools. |
Thesis: Agent marketplaces are the “app store” moment. Fetch.ai’s 2.7M agents shows the demand. The winner will be the one that solves trust/reputation for agents.
9. Agent-Native Voice Infrastructure
Voice as an agent interface, not a human IVR system.
| Company | Funding | What it does | Performance |
|---|---|---|---|
| Vapi | $50M Series B | Voice AI infrastructure. API-first, agents are primary entity. | 1B calls, 2.5M+ agents, sub-500ms |
| Bland AI | — | Owns full stack (LLM + STT + TTS + telephony). Millions of autonomous calls. | 400ms latency (vs 1240ms industry) |
| LiveKit Agents | ⭐ 11k | Open-source real-time voice/video agents via WebRTC. | Multi-agent handoff built-in |
Thesis: Voice agents are the first agent-native consumer product at scale. Customer service, sales, scheduling — all being replaced by autonomous voice agents.
10. Agent-Native UX (Agent-Human Handoff)
The interface between agents and the humans who oversee them.
| Company | Funding | What it does |
|---|---|---|
| CopilotKit + AG-UI | $27M Series A | Agents generate their own UI in real-time. Bi-directional protocol. Adopted by Google, AWS, MS. |
| Gumloop | $50M Series B (Benchmark) | Visual canvas for multi-agent orchestration. Agents operate autonomously on schedules. |
Thesis: Chat is not the right UX for agents. Generative UI (agents create their own interface) and visual orchestration canvases are the agent-native interaction patterns.
11. Agent-Native Execution & Sandboxing
Where agent code actually runs.
| Company | Funding/Stars | What it does |
|---|---|---|
| E2B | $21M Series A | Firecracker microVMs for agent code execution. <200ms boot. Desktop Sandbox for computer-use agents. |
| Daytona | ⭐ 73k | Sub-90ms sandbox spin-up. Snapshots/restore. MCP server built-in. |
| Inngest AgentKit | ⭐ 5k | Durable execution for agent workflows. Built for LLM unreliability (retries, fallbacks, throttling). |
Major Platform Moves (Agent-Native Pivots)
These aren’t startups — they’re billion-dollar companies redesigning for agents:
| Company | What they’re doing |
|---|---|
| Stripe | Agent Toolkit + Link as agent wallet |
| Shopify | Storefronts for AI shopping agents |
| Robinhood | Agents trade stocks directly |
| Visa/Mastercard | Backing Skyfire for agent payments |
| Notion | Workspace as agent hub |
| Atlassian | Third-party agents in Confluence |
| WordPress.com | AI agents write and publish |
| Palo Alto Networks | Acquired Portkey (agent security gateway) |
| Snyk | Acquired Invariant Labs (agent security) |
The Agent-Native Stack (Emerging)
┌─────────────────────────────────────────────────────────────────┐
│ AGENT-NATIVE STACK │
│ │
│ ┌─────────────────────────────────────────────────────────────┐│
│ │ PROTOCOLS: MCP (tools) + A2A (agents) + AG-UI (humans) ││
│ └─────────────────────────────────────────────────────────────┘│
│ │
│ ┌──────────┐ ┌──────────┐ ┌──────────┐ ┌──────────────┐ │
│ │ Identity │ │ Payments │ │ Security │ │ Observability │ │
│ │ Skyfire │ │ Stripe │ │ Credal │ │ AgentOps │ │
│ │ KYA │ │ Skyfire │ │ Arcjet │ │ Braintrust │ │
│ └──────────┘ └──────────┘ └──────────┘ └──────────────┘ │
│ │
│ ┌──────────┐ ┌──────────┐ ┌──────────┐ ┌──────────────┐ │
│ │ Search │ │ Browser │ │ Memory │ │ Execution │ │
│ │ Exa │ │ Browser- │ │ Letta │ │ E2B │ │
│ │ Tavily │ │ base │ │ Mem0 │ │ Daytona │ │
│ └──────────┘ └──────────┘ └──────────┘ └──────────────┘ │
│ │
│ ┌─────────────────────────────────────────────────────────────┐│
│ │ MARKETPLACES: Fetch.ai (2.7M agents) + Smithery + Composio ││
│ └─────────────────────────────────────────────────────────────┘│
│ │
│ ┌─────────────────────────────────────────────────────────────┐│
│ │ UX: CopilotKit (generative UI) + Voice (Vapi, Bland) ││
│ └─────────────────────────────────────────────────────────────┘│
└─────────────────────────────────────────────────────────────────┘Key Funding Signals
| Company | Amount | Round | Category |
|---|---|---|---|
| Exa | $250M | Series C | Agent search |
| Gumloop | $50M | Series B | Agent building |
| Vapi | $50M | Series B | Voice agents |
| CopilotKit | $27M | Series A | Agent UX |
| Tavily | $25M | Series A | Agent search |
| E2B | $21M | Series A | Agent execution |
| AgentOps | $2.6M | Seed | Agent observability |
| Portkey | Acquired | Palo Alto Networks | Agent security |
| Invariant | Acquired | Snyk | Agent security |
Emerging Thesis: Where Value Accrues
- Identity + Payments = the unlock for agent autonomy. Without them, agents always need a human in the loop.
- Web access = the biggest bottleneck today. 85% of the web has no API. Agent-native search ($250M Exa) and browser infra (Browserbase) are the bridges.
- Security + Governance = the enterprise gate. Acquisitions by Palo Alto and Snyk confirm this is where enterprise money flows.
- Self-improvement = the product differentiator. Hermes (190k stars) proved users want agents that learn. Static agents will lose to learning agents.
- Marketplaces = the network effect play. Fetch.ai (2.7M agents) is early but shows the direction — agents hiring agents creates exponential value.
Research Questions for Further Posts
- How does agent-native security differ from traditional zero-trust?
- What does an “agent-native database” look like? (Read/write patterns, schema design)
- How do agent reputation systems work? (Trust without human reviews)
- What’s the business model for agent marketplaces?
- When does agent-native replace human-first? (Adoption curves, triggers)
- What’s the regulatory landscape for autonomous agent transactions?