Stateless and Blind: Two Gaps Every Agent Eventually Hits
TinyFish and Hermes Agent represent the two gaps most agent stacks hit: real-time web access and persistent cross-session learning.
Most agents are stateless and blind. They start from scratch on every session, and they have no access to the live web unless you build that access yourself. These aren't implementation details. They're architectural constraints that surface sooner or later in every serious agent project.
Two tools I've been looking at sit at different parts of this problem.
TinyFish
TinyFish is an infrastructure platform for agents that need to interact with the real web. Four APIs: Search, Fetch, Agent, and Browser. Search renders pages in live browsers to return fresh structured data, not cached results. Fetch converts rendered pages into clean markdown or JSON for LLM consumption. The Agent API automates multi-step workflows including form filling and authenticated logins, with 89.9% accuracy on Mind2Web benchmarks. Browser provides stealth Chromium sessions that bypass anti-bot detection with sub-250ms cold starts.
It's MCP-native. Works with Claude and Cursor out of the box. Single API key, unified credit pool, Search and Fetch are free.
The problem TinyFish solves isn't scraping. Agents operating on real tasks, competitive intelligence, compliance monitoring, vendor data, keep hitting walls where the data they need is dynamic, authenticated, or behind bot protection. TinyFish is infrastructure for agents that need to see the web as it is, not as it was cached six hours ago.
Hermes Agent
Hermes, from Nous Research, is an autonomous agent built around a closed learning loop. The system creates skills from experience, refines them during operation, and builds persistent knowledge across sessions. Not a one-shot tool. An agent that gets better at the work the more it does.
It runs on local machines, Docker, SSH, or serverless infrastructure (Daytona, Modal). It connects to 20+ messaging services from a unified architecture. 60+ built-in tools plus MCP server integration. Skills are portable, compatible with the agentskills.io standard.
The interesting part is what Hermes does with the session boundary. Most agents treat sessions as atomic: start, task, end. Hermes treats sessions as data. What the agent does becomes training material for what it does next. That's a different operating model.
The Signal
TinyFish addresses perception: can the agent access the information it needs? Hermes addresses continuity: does the agent improve from doing it?
Both gaps usually get papered over with prompt engineering. They don't go away. The project just gets serious enough that they can't be ignored anymore.
The move worth watching is from stateless, isolated agents toward persistent, web-aware systems that improve through use. Neither tool alone closes the loop. Together, they sketch what that architecture starts to look like.