Agents vs. workflows: choosing the right pattern in 2026

The word "agent" did a lot of work between 2023 and 2026, and most of it was marketing. If you sit through enough sales decks, you come away thinking every LLM call should be wrapped in a multi-step planner with tool access and a name like Atlas or Sage. In practice, the most reliable systems are still the simplest ones — and a lot of what gets shipped as an "agent" is a workflow in a costume.

This post is about the actual distinction, when each pattern earns its keep, and how to choose without getting talked into more complexity than your problem needs.

The architectural difference, not the marketing one

The cleanest definition came from Anthropic in their December 2024 guide on building effective agents. At Anthropic, they categorise both as agentic systems, but draw an architectural distinction: workflows are systems where LLMs and tools are orchestrated through predefined code paths, while agents are systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks.

That is the whole distinction. A workflow has a deterministic graph that you wrote. An agent has a loop where the model decides what to do next.

The same Anthropic post — which is still the best primer on this topic eighteen months later — went on to describe five composable workflow patterns: prompt chaining, routing, parallelisation, orchestrator-workers, and evaluator-optimiser. None of them are "agentic" in the autonomous sense. All of them are useful. Most of them are what teams actually need.

Anthropic's own recommendation on when to use which was unambiguous. When building applications with LLMs, find the simplest solution possible, and only increase complexity when needed. The follow-up was even more pointed: success in the LLM space is not about building the most sophisticated system — start with simple prompts, optimise them with comprehensive evaluation, and add multi-step agentic systems only when simpler solutions fall short.

A year and a half later, this advice has aged extremely well.

Why workflows usually win

If you actually instrument production systems, the case for workflows is hard to argue with on most problems.

Predictable cost. A workflow has a known number of LLM calls. A workflow can handle thousands of requests per minute, at predictable costs and latency, with zero risk of runaway behaviour. An agent runs until it decides it is done. When something goes wrong — and it will — you do not get a nice red error message, you get a token bill that looks like someone mistyped a loop condition and summoned the API 600 times.

Predictable latency. Same logic. A four-step workflow takes roughly the time of four LLM calls plus your tool calls. An agent takes between one and however-many-iterations-it-decided-it-needed-this-time. AI workflows are best for structured, repeatable tasks — their main strength is predictability, because the execution path is fixed, costs and latency are consistent, and debugging is straightforward.

Debuggability. A workflow is just code with an LLM in the middle. You can put a breakpoint on it. You can replay it. You can write a unit test for it. If you are working in healthcare, finance, law, or government — places where "we think the AI decided to try something new" is not an acceptable answer — this matters. You cannot build a safe AI system without transparency, and workflows give you that by default.

Compliance fit. This is going to become structural rather than optional. The EU AI Act's obligations for high-risk systems land on 2 August 2026 with documentation and human-oversight requirements that a deterministic pipeline can satisfy more cheaply than an autonomous agent ever will.

The case for workflows can be summarised as: when the steps are knowable, encoding them is cheaper, faster, and easier to operate than asking a model to figure them out at runtime.

Where agents actually earn their keep

There is a real category of problems where the workflow case breaks down, and it is worth being precise about it. Agents excel in scenarios requiring flexibility and large-scale decision-making — when the steps are genuinely unpredictable and the value of getting them right outweighs the cost of variable execution.

The honest list, in our experience:

• Investigation and exploration. Tasks where the next step depends on what the last step uncovered — debugging a system, researching a topic, triaging a support ticket across multiple systems.
• Variable-shape inputs. When the same task arrives in materially different forms each time, a fixed pipeline ends up branching so many times that the agent loop becomes simpler than the workflow it replaces.
• Coding agents. This is the use case where agents have demonstrably worked. Anthropic describes their own coding agents using an orchestrator-workers pattern to handle GitHub issues across multiple files without predefined subtasks. The structure of "explore the codebase, propose a change, run the tests, react to failures" maps cleanly to an agent loop.
• Long-horizon tasks with tool use. Anything where the model genuinely needs to chain twenty or thirty tool calls based on intermediate results.

Notice what is not on the list: most of what gets sold as agentic in 2026. Customer support routing, document classification, content extraction, RAG-with-a-reranker, "agentic" sales follow-up — all of these are workflows. Treating them as agents is over-engineering. Many so-called agentic RAG designs are actually workflows, not agents — and they tend to work better when teams stop pretending otherwise.

The framework landscape, briefly

You can build either pattern with no framework at all. For workflows this is often the right answer; ten lines of Python orchestrating two or three LLM calls is more maintainable than a graph DSL you also have to learn.

If you do reach for a framework, the 2026 landscape has settled somewhat. LangGraph has the fastest-growing adoption with 45,000+ GitHub stars as of early 2026, and is the most production-mature option with enterprise support contracts, versioned APIs and backward compatibility guarantees. AutoGen split into Microsoft's v0.4 rewrite and the community-continued AG2 lineage, which has caused real confusion — pick one deliberately. CrewAI has matured as the gentlest learning curve at the cost of less control. OpenAI's Agents SDK reached production maturity for OpenAI-native deployments. Anthropic shipped a Claude Agent SDK that has gained traction for production deployments because of native tool use and memory features.

The most important thing to internalise about frameworks is Anthropic's own warning, which still applies: if frameworks are used, a deep understanding of the underlying code is crucial to avoid errors due to incorrect assumptions. We have lost count of the number of times a team has burned three weeks chasing a bug that turned out to be an opinionated default inside their agent framework. Frameworks make starting easier and debugging harder. Trade accordingly.

How to choose

The decision process we use with clients, in roughly this order:

1. Can a single well-prompted LLM call do this? If yes, do that. Add retrieval if needed. Stop. For many applications, optimising single LLM calls with retrieval and in-context examples suffices.

2. Can a deterministic workflow do this? If you can sketch the steps on a whiteboard, encode them. Use routing, chaining, parallelisation as needed. Almost every system we put into production lives here.

3. Does this task genuinely have an unpredictable shape? If the answer is yes — meaning, you cannot enumerate the branches in advance without missing real cases — then an agent loop earns its place. Even then, structure the agent narrowly: small tool set, clear stop conditions, explicit budget on iterations and tokens, observable trace of every step.

4. Will a human be in the loop on irreversible actions? Agents that can spend money, send emails, or modify production systems need a human approval gate. The frameworks that lead on production maturity — LangGraph, the OpenAI and Claude SDKs — make human-in-the-loop checkpoints first-class. Use them.

The reason this order matters is that complexity compounds. An agent on top of a workflow that you only half-needed produces a system that nobody can debug at 2am when it is making the wrong call in front of a paying customer.

What we tell teams in 2026

A few opinionated rules of thumb that have held up across the last year of client work:

• Default to workflows. If you find yourself reaching for an agent, write down why a workflow cannot do it before you proceed. Half the time the act of writing it down resolves the question.
• Bound your agents. Maximum tool calls per task. Maximum tokens per task. Maximum wall-clock time. Hard fail when exceeded. Without these you will burn money on confused loops.
• Make the trace observable. Every tool call, every model response, every decision point logged with structured data, not free-text. You will need this the first time the agent produces a wrong answer in front of a real user.
• Evaluate the whole trajectory, not just the final answer. A correct final answer reached via an absurd path is a system that will fail unpredictably tomorrow. Evaluate the path.
• Be honest with yourself about what is autonomous. A "multi-agent system" with three LLM calls and a router is two prompts and a switch statement. Calling it agentic does not change what it is. The agent-framework landscape has converged on common abstractions, with differentiation in ecosystem depth — picking the right tool for the actual shape of the problem matters more than picking the most powerful tool available.

The pattern we see win, consistently, is teams who treat the workflow-or-agent decision as an architecture question rather than a status one. Workflows are not less impressive. They are more reliable, cheaper to operate, easier to certify, and they ship sooner. Agents are not more advanced. They are appropriate for a specific shape of problem and a nightmare everywhere else. Treating both as tools, not identities, is what separates the teams shipping production AI from the teams demoing it.

References

1. Anthropic — Building effective agents. https://www.anthropic.com/news/building-effective-agents
2. Towards Data Science — A Developer's Guide to Building Scalable AI: Workflows vs Agents. https://towardsdatascience.com/a-developers-guide-to-building-scalable-ai-workflows-vs-agents/
3. Decoding AI — AI Workflows vs Agents: The Autonomy Slider. https://www.decodingai.com/p/ai-workflows-vs-agents-the-autonomy
4. Towards AI / Praveen Subramanian — Agent Workflow Patterns: Beyond Anthropic's Playbook. https://pub.towardsai.net/agent-workflow-patterns-beyond-anthropics-playbook-1bd76a48d63d
5. Pharos Production — AI Agent Frameworks 2026: Comparison Guide. https://pharosproduction.com/insights/engineering/ai-agent-frameworks-comparison-2026/
6. Prompt Engineer Collective — AI Agent Frameworks Compared: LangGraph vs CrewAI vs AutoGen (2026). https://pecollective.com/blog/ai-agent-frameworks-compared/
7. Alice Labs — AI Agent Frameworks 2026: Production-Tested Ranking. https://alicelabs.ai/en/insights/best-ai-agent-frameworks-2026
8. AI Exploration Journey — Building LLM-based Agents from the Toolbox. https://aiexpjourney.substack.com/p/building-llm-based-agents-from-toolbox
9. EU AI Act Service Desk — Timeline for the Implementation of the EU AI Act. https://ai-act-service-desk.ec.europa.eu/en/ai-act/timeline/timeline-implementation-eu-ai-act

Written by JP, Sixees Labs. Last reviewed May 2026.