Issue 013 organizations name itnewest evidence Dec 2024
An agent is being built where a predefined workflow would do the job
The architectural distinction is a decision about who chooses the next step - your code, or the model. Most deployments hand that choice to the model without having asked whether they needed to.
Anthropic states the trade explicitly: agents buy task performance with latency better task performance, and you should consider when that trade makes sense. Where the steps are knowable in advance, a workflow gives you the same outcome with behaviour you can test, cost you can predict, and failures you can reproduce. The cost of getting this wrong is not a bad answer - it is a system whose behaviour you cannot characterise, attached to a problem that never required it.
How to fix it — 2 approaches, 5 steps
Decide who chooses the next step, in writing
For each candidate, state whether the sequence of steps is knowable in advance. If it is, the answer is a workflow, and the decision is made.
Done when Each proposed agent has a written answer to whether its steps are knowable in advance, anything with a knowable sequence was built as a workflow, and the latency and cost premium accepted for dynamic control is recorded.
- For each proposed agent, write down the steps. If you can write them, you do not need an agent.
- Reserve dynamic control for cases with genuine exceptions and open-ended paths.
- Record the latency and cost premium accepted in exchange, so the trade is explicit.
Reach for the composable patterns before autonomy
Prompt chaining, routing and parallelization solve a large share of real cases with predictable behaviour and no autonomous control loop.
Done when The record for each autonomous loop in production states which deterministic pattern was tried first and how it demonstrably failed.
- Try the deterministic pattern first: chain, route, or run in parallel and combine.
- Escalate to an autonomous loop only where the pattern demonstrably fails.
The evidence — 4 documents
| Organization | Document | Position |
|---|---|---|
| AnthropicFrontier lab · December 2024 | Building effective agentsOur reading Draws the architectural distinction - workflows orchestrate models and tools through predefined code paths, agents dynamically direct their own processes and tool use - and recommends finding the simplest solution possible, increasing complexity only when needed, which may mean not building an agentic system at all.Section: workflows versus agents; when (and when not) to use agents | names it |
| IBMHyperscaler | Agentic AI risk and opportunityOur reading Names the choice outright rather than assuming the agentic route: it holds that the simplest answer to most business problems is a designed orchestration of conventional models and language models, that agents will not suit every problem, and that an agent is warranted only where the complexity of the problem earns it - because the harder architecture behaves less predictably.Risks and key mitigations of autonomous action | names it |
| OpenAIFrontier lab | Practical guide to building agentsOur reading Defines a workflow as a sequence of steps that must be executed to meet a goal, and an agent as a system that independently accomplishes tasks on the user's behalf, reserving agents for workflows involving nuanced judgment, exceptions and context-sensitive decisions that have resisted conventional automation.What is an agent?; when should you build an agent? | names it |
| OpenAIFrontier lab | Practical guide to building agentsOur reading Contrasts a traditional rules engine working like a checklist against an agent evaluating context and subtle patterns, positioning the agent as the answer only where clear-cut rules do not apply.When should you build an agent? | names it |