Agents
A practical field guide to what "agentic" means once AI systems stop merely answering and start acting.
Signal summary
Chip Huyen's essay is useful because it strips away the vague excitement around agents and turns the topic into a set of design questions: what tools can the model use, what plans can it make, what feedback does it receive, what state does it keep, and how do we know whether the workflow worked? The signal is not "agents are magic." The signal is that agentic systems are becoming a new production design discipline.
Why it matters
For LGI, this is a foundation stone. The serious question is no longer whether an LLM can produce a plausible next step. It is whether a system can structure action, memory, review, and escalation around the model. Agents become useful when they are embedded in a workflow that knows what evidence is available, what tools are allowed, what success means, and where human judgement enters.
LGI reading
Agentic work is not just a model capability. It is an operating model. The practical value lives in the scaffolding: source context, task decomposition, tool boundaries, evaluation loops, and recovery paths. This is why LGI frames agent work through knowledge systems and governance rather than prompt tricks.
Use this if you are thinking about
AI-assisted delivery workflows, coding agents, research agents, internal knowledge assistants, human-in-the-loop operations, and the move from chat interfaces to action systems.
Related LGI concept
Agentic Workflows
Related signals
Paper · Agentic Workflows
Drop the Hierarchy and Roles
Argues self-organising agents can outperform designed hierarchies under some conditions.
Paper · Agentic Workflows
Building Persona-Based Agents On Demand
Dynamic agent personas tailored to user/task context instead of static workflows.