Monokit is on sale: the app screens most kits leave out · See Monokit → · Read the post →

← Studio

Engineering · 2026

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 every session from scratch, and they can't see the live web unless you build that access yourself.

I know the first one well. My audit agent, Ledger, has no memory of yesterday's run, which is both why it's reliable and why it repeats itself. Neither gap is an implementation detail. Sooner or later, every serious agent project runs into both.

Two tools I've been reading about go after one gap each.

TinyFish: the web as it is now

TinyFish is infrastructure for agents that need the real web. It has four APIs:

APIWhat it does
SearchRenders pages in live browsers and returns fresh structured data, not cached results
FetchTurns rendered pages into clean markdown or JSON for a model
AgentAutomates multi-step workflows, including forms and authenticated logins (89.9% on the Mind2Web benchmark)
BrowserStealth Chromium sessions that get past bot detection, with sub-250ms cold starts

It's MCP-native, works with Claude and Cursor, and runs on one key and one credit pool. Search and Fetch are free.

Scraping isn't the interesting part. Agents doing real work (competitive intelligence, compliance monitoring, vendor data) keep hitting pages that are dynamic, behind a login or behind bot protection. TinyFish is for agents that need the web as it is now, not as it was cached six hours ago.

Hermes: learning from the last session

Hermes, from Nous Research, is built around a closed learning loop. It writes skills from experience, refines them as it works and keeps what it learns across sessions.

It runs locally, in Docker, over SSH or serverless (Daytona, Modal), talks to 20+ messaging services, and ships 60+ tools plus MCP support. Its skills follow the agentskills.io standard, so they travel.

What caught my eye is the session boundary. Most agents treat a session as atomic: start, task, end. Hermes treats it as data, so what it did today becomes material for what it does tomorrow. That's a different operating model from anything I run.

What I take from it

TinyFish is about perception: can the agent get the information it needs? Hermes is about continuity: does it get better at the job?

Both gaps usually get papered over with prompt engineering. That works until a project gets serious enough that it doesn't. Neither tool closes the loop on its own, but together they sketch where agents are heading: away from stateless one-shots, towards systems that see the live web and improve through use.

Ledger is the case in point. A memory of what it had already reported would have saved me fourteen duplicate issues. It would also have made it a little easier to talk itself out of the fifteenth.