Loading lesson…
Applied LLM Systems: RAG, Agents & MCP / MCP and Workflows
Deterministic, agentic, autonomous.
Reviewed by Yuvaraj
"Agent" gets attached to almost anything today, which hides a useful distinction. Some systems follow fixed steps; some decide their own steps; a few run on their own for long stretches. Knowing which you are building, and which you actually need, is one of the most practical judgments in applied AI.
Think of a line from fully predetermined to fully self-directed:
The difference between the first two is who decides the order of steps, a developer ahead of time, or the model in the moment.
The instinct to reach for an agent is often wrong. Prefer the simplest thing that works:
Answer from memory before revealing, retrieval practice is what builds durable recall.
How does a deterministic workflow differ from an agent?
Ask about this lesson, or about anything in AI. Answers cite the lessons they draw on.
Finished this lesson?
Mark it complete to earn XP, keep your streak, and schedule a review.
Reach for the loop last
A widely echoed piece of practical advice: find the simplest pattern that solves the problem, and add agentic decision-making only when a fixed workflow genuinely cannot express the task. Autonomy is a cost, not a feature.
The further right on the spectrum, the more an error can compound before anyone notices. A wrong step in a fixed workflow is contained by the next fixed step; a wrong step in a long-running autonomous agent can lead to more wrong steps. This is why autonomous systems need the strongest guardrails: step and cost budgets, scoped permissions, checkpoints, observability, and human approval for consequential actions.
Label the confidence
Fully autonomous, reliably self-improving systems are an area of active research, not a solved capability. Today's dependable deployments are overwhelmingly workflows and tightly-scoped agents with humans in the loop. Treat stronger claims as emerging or speculative, not established.