Stop Building AI Agents the Old Way

Defining clear success criteria in spec docs for a long-running agent as a contract rather than a loose prompt is a major W. You must design robust evaluation harnesses with deterministic verifiers to catch model drift. Multi-agent systems often suffer from context rot during long loops. You need session mining to convert past failures or success into rules but rules are ignored too. Integrating observability tools and HITL is non-negotiable for real-world agentic reliability, mic drop!

the point about treating spec docs like a contract instead of a loose prompt is so real, most agent failures i've dealt with trace back to vague success criteria not the model. we ran into the same context loss issue on a long-running agent and ended up using HydraDB to keep persistent state across sessions so it stopped re-deriving everything from scratch each run.

I wonder about context, no ome mentioned if it is compacted several times, is there a way to handoff and stsrt another session in the loop? I think this would be more efficient.

what are the most important ones for a beginner? can you please create some follow up videos on the topic?

Architecture makes perfect sense—now show us the stack! Are you using LangGraph or custom Python state machines for the outer loop?

For once, seeing a full example first of this brick wall of very specific architecture would be cool. Because that's a whole lot of words, my friend.

It's impressive that it explained a lot what Hermes agent has been working on

👍👍👍

is there any resources to learn how learn build long running agent in code. rn I am trying to explore by myself in langgraph, but not find any clear path

Great breakdown of the 7-component architecture! You covered the theory of the outer loop, evaluators, and verifiers perfectly, but I'm curious about your actual tech stack. Are you building these long-running agents using a framework like LangGraph/CrewAI or are you rolling your own state machines in Python? Also, what are you using for the deterministic verifier steps (Pydantic, E2B)? Would love a follow-up video showing the actual tools/repo/code implementation of this loop!