Keep source evidence through every stage

An original publishing workflow could collect source facts, draft a short explanation, then check that every factual claim has support. Keep the source notes available at the final check instead of relying only on the draft.

Keep only stages that improve verification

Three calls are not automatically better than one. The extra stages add delay and can repeat the same error.

Keep a stage only if it catches failures or makes a difficult task easier to verify.

Why a chain is easier to debug

A single prompt that says “research, compare, and recommend” hides the failure. You see a polished ending and cannot tell which part invented a source.

A chain makes the intermediate text visible. You can reject a bad extract before a later step writes a recommendation on top of it.

Each step should have a narrow output. “List the dates mentioned, and say if a date is missing” is checkable.

“Understand the contract” is not. Pass the structured output forward, not a rambling paragraph the next model has to reinterpret.

If step two needs a field, step one should be required to produce that field or an explicit empty value.

Where chains waste a call

Every step pays for the tokens it reads and writes. A chain that restates the whole document at every stage is an expensive copy.

Pass the smallest artifact that the next step needs. Also avoid a chain whose only purpose is to look like a pipeline.

If one prompt already produces a checkable result, keep one prompt.

Decide the failure behavior. If the extractor returns no date, the summarizer should not guess one.

Code should branch. A chain that always continues will launder an empty step into a confident final answer.

That is worse than a single prompt, because the final answer looks like it survived review.

What to record before you rely on it

Draw the steps and the artifact each one must emit. Save one run where step one failed and confirm that later steps did not run.

Record the token cost of the whole chain on a real document, not on a one-paragraph example.

Keep the model id, the date, and the inputs you used beside that note, so a later change can be compared with the same test. When the result is a failure, store the artifact and change the design.

Do not relax the check until the model happens to pass. Name a person who owns that failure.

A note with no owner and no inputs cannot be repeated, and a pass with no date cannot be compared with the next model id.

Common questions

Is a chain the same as an agent?

No. In a chain, you chose the steps. An agent chooses them at runtime.

A chain can still call tools, but the order is yours.

How many steps is too many?

When a step does not have a check, or when it only rewrites the previous text. Add a step only if its output can fail a test the previous output could not.

Sources & further reading

  1. Anthropic: Building effective agents ↗
How this story was made

Written by Kristian Kostov with AI assistance and checked against the linked sources. Company performance claims are attributed to the company. Analysis reflects AiLookout’s interpretation; we have not independently tested the products discussed. Cover photography is illustrative and does not depict the specific announcement or product.

Our editorial standards
Back to all stories