Connected AI agents: when they make sense
What separates a workflow from an agent, the most common patterns for agents working together and why the simplest solution is often the best one.
Related project: ETUNIM AI Studio
There is a lot of talk about AI agents, often without explaining what they are. This article summarises the key concepts, based on a technical guide published by Anthropic in December 2024.
Workflow or agent?
The distinction is useful:
- Workflow: AI models and tools follow a path defined in advance.
- Agent: the model itself decides which steps and tools to use, based on the result of each action.
A workflow is more predictable. An agent adapts better to open-ended problems, but can cost more and compound errors.
Ways to connect them
The guide describes several patterns:
- Chaining: each step uses the output of the previous one, with checks along the way.
- Routing: the request is classified and sent to the most suitable task.
- Parallelisation: several tasks run at the same time and the results are combined at the end.
- Orchestrator and workers: a central model splits the work, hands it out and merges the answers. The subtasks are not defined in advance.
- Evaluator and optimiser: one model produces and another evaluates, in a loop. It works when the evaluation criteria are clear.
More agents is not always better
The advice is to start with the simplest solution and only add complexity when needed. Often, a single well-prepared call is enough.
At ETUNIM, too, we see connecting agents as a means, not a goal. It makes sense when dividing the work makes the result better or easier to check.
What this means for us
ETUNIM AI Studio, still at the concept stage, starts from these ideas: agents that collaborate, share tasks and recognise errors, with visible steps and the user in control.
Source
- Anthropic, Building effective agents, December 2024.