Predefined path or dynamic direction: which steps deserve an agent

The Tatano Energy platform serves 7 languages across 4 country domains, publishes 8 SEO articles every day and requires 0 manual intervention. Described that way, it sounds like a system where a model decides what to do next. Most of the steps in a pipeline like that should not work that way. Anthropic draws the line in a usable place, and it is worth applying step by step rather than to the system as a whole.

The line the definition draws

Workflows, in Anthropic's framing, are systems where LLMs and tools are orchestrated through predefined code paths. Agents are systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks. The difference is not how clever the model is, or how many tokens the step consumes. It is who chooses the next call: your code, or the model.

That makes classification a question you can answer by reading your own code. If the sequence of tool calls is the same on every run and only the content changes, the step is a predefined path with a language model inside it. If the model can pick a different tool, or call the same one a different number of times, you have delegated control and inherited everything that comes with it.

Count the decisions, not the prompts

Walk a daily publishing pipeline and most steps turn out to have no runtime decision in them. Selecting the next topic from a trends feed, producing language variants, attaching structured data, pushing to the right country domain: each has a fixed shape. A model writes text inside those steps, but the order is fixed by code. Calling the whole thing an agent obscures which single step actually directs itself.

The practical payoff is scoping. Steps on a predefined path fail in ways you can enumerate and test before launch. Steps where the model directs itself fail in ways you tend to discover in production. Knowing which is which tells you where to concentrate guardrail effort, instead of spreading it evenly across a pipeline where most stages do not need it.

Dynamic direction is not free

Anthropic's own measurements give the price of the choice. In their data, agents typically use about 4× more tokens than chat interactions, and multi-agent systems about 15× more tokens than chats. That is the visible cost. The quieter one is stated alongside it: the autonomous nature of agents means higher costs and the potential for compounding errors, which is why they recommend extensive testing in sandboxed environments along with appropriate guardrails.

Compounding errors matter more than the bill when nothing is supervised. A fixed step that produces something wrong produces one wrong thing. A step that directs itself can take a wrong intermediate result and build further calls on top of it before anything returns. At 8 publications a day with 0 manual intervention, no reviewer is standing between that chain and the live site.

Where agency earns its place

Recovery is the honest case for it. The Matthieu Pesesse platform publishes articles autonomously on a daily basis with 3 model fallback levels and roughly 0 hours of weekly content operations. Choosing a level depends on what just happened, so something has to make that call at runtime. It is a narrow, bounded delegation: the choice is which generator to use, not which tools to invoke next.

Keep that shape wherever you can. Let code own the sequence and let the model own one choice inside it, from a list you wrote. The step stays testable because the set of possible outcomes is still finite, and the blast radius of a bad decision is one substituted component rather than an open-ended chain of calls nobody predicted.

Guardrails follow the classification

A cap on iterations only applies to a step that can iterate. Anthropic notes it is common to include stopping conditions, such as a maximum number of iterations, to maintain control. Applied to the steps you classified as self-directing, that does real work. Applied to a fixed sequence, you have added a counter that never trips and a configuration value someone will later assume is protecting something.

The other half is deterministic. Anthropic describes combining the adaptability of agents with deterministic safeguards like retry logic and regular checkpoints, and building systems that can resume from where the agent was when the errors occurred. They also had agents summarise completed work phases and store essential information in external memory before proceeding. Checkpoint boundaries are easiest to place where code owns the sequence, which is another reason to keep the self-directing part small.

Before building, walk the pipeline step by step and write down who picks the next tool call. Where your code picks, enumerate the failure modes and test them. Where the model picks, add stopping conditions, checkpoints and sandboxed testing before anything runs unsupervised. The useful output is not a label for the whole system. It is a shorter list of places that need the expensive treatment.

Sources

Anthropic — Building effective agents — https://www.anthropic.com/engineering/building-effective-agents

Anthropic — How we built our multi-agent research system — https://www.anthropic.com/engineering/multi-agent-research-system

Neurolinks case study — Four markets, one codebase — https://neurolinks.be/work/tatano-energy

Neurolinks case study — A personal brand that publishes itself — https://neurolinks.be/work/matthieu-pesesse-media