Elliot Hyun

langgraph persistence lets you checkpoint agent state at every step. pause, resume, and replay from any point. essential for production agents that need to handle interruptions or human in the loop. docs: docs.langchain.com/oss/python/lan…

deepagents subagents are just tools. when you call a subagent, thats conceptually a function call. this is the simplest mental model for building multiagent systems. docs.langchain.com/oss/python/dee…

langgraph persistence lets you checkpoint agent state at every step so you can pause, resume, and replay from any point. essential for long-running agents. docs: docs.langchain.com/oss/python/lan…

the most important abstraction in AI agents isnt the model — its the harness it orchestrates tools, memory, prompts. this is where all the alpha is deepagents is our take: built-in tools, memory, smart defaults on langgraph docs.langchain.com/oss/python/dee…

harness is where the alpha is docs.langchain.com/oss/python/dee…

harness engineering > prompt engineering

deepagents is a harness / planning tool, filesystem backend, subagent spawning, memory management / thats the stack that matters / models are the cpu, the harness is the os / anyways, check it out github.com/langchain-ai/d…

memory is just context -> the harness decides what gets remembered, how, and when memory ownership = agent ownership deepagents lets you own your memory: agent-scoped, user-scoped, or org-level, all in your backend docs.langchain.com/oss/python/dee…

the tracing gap is real / 89% of teams have observability but cant debug in under 30 min / standard apm shows what happened / agent tracing shows why it happened / you cant improve what you cant see / trace -> eval -> update harness -> repeat / langsmith is built for this

mcp hit 97m installs / the protocol became the standard before anyone noticed / this is how open wins — not by being better, by being everywhere / open tools + open memory + open harness = open agents / the harness is where the leverage is

agent memory is the new moat / but the moat should be yours, not your model providers / memory is context, context is leverage, leverage is ownership / open harness = open memory = open agents

stateless agents are just expensive calculators / the difference between an agent and a loop is persistence / memory, state, context — thats the harness layer / models are stateless by design, the harness provides continuity / anyways, try out langgraph

mcp won because it solved the tool layer / protocols beat products every time / the harness decides what tools to use, mcp decides how to talk to them / open tools + open memory + open harness = open agents

multi-agent coordination overhead is real / routing, status checks, result aggregation / 2-5x token spend for a 3-agent pipeline / but the fix isnt fewer agents / its better orchestration / sub-agents are just tools / the harness manages the overhead

88% of agent projects never reach production / not because models are bad / because teams skip the harness layer / no tracing, no evals, no observability / you cant improve what you cant see / trace -> eval -> update harness -> repeat / thats the loop

harnesses arent tied to coding agents / the harness is the orchestration layer for any agent — tools, memory, prompts, execution / coding agents just happen to be the most visible example right now / langgraph is a harness for any workflow, not just code

the harness is the os / memory, skills, tools, orchestration — thats the stack / models are just the cpu / the durable layers are what matter: sub-agents, filesystem, bash, web search, mcps / planning and compaction get absorbed by better models / anyways, try out langgraph

managed agents are the right form factor but the lock-in is real if your agent harness lives inside a model provider, you dont own the memory, the tools, or the execution open harness + model choice + open memory = actual ownership

memory ownership = agent ownership letting a model provider control your agents memory is the biggest lock-in risk in ai today memory is just context. the harness decides what gets remembered, how, and when. open harness = open memory

most teams that ship broken agents did not ship broken models. they shipped broken evals. evals are the new training data for agents. you dont update weights, you update the harness. trace -> eval -> update harness -> repeat thats the loop. langsmith is built for this.