Sources checked
Reconcile worker results instead of concatenating them
An original document comparison can give each worker one source and a common evidence format. The coordinator should reconcile conflicting definitions and duplicate claims, rather than just concatenate the outputs.
More workers can repeat the same assumption
More agents can multiply the same mistaken assumption. Compare against a single-worker baseline and inspect the disagreements that the final result hides.
Where the extra tokens go
Each agent has a prompt, a transcript, and tool results. When one agent messages another, that message is output for the sender and input for the receiver.
A debate of five rounds is five copies of the growing argument. The bill is not “one hard question.”
It is the protocol you invented between the roles.
Failures multiply. If agent A’s plan is wrong, agent B may still spend a full context implementing it, and agent C may spend another context reviewing the wrong implementation.
A single model with a checklist would have failed once. Write down that cascade before you call the design more reliable.
When a second role pays for itself
A second role helps when the check is real and cheaper than the work. A reviewer that only sees a diff and a test result can be smaller than the coder.
A reviewer that rereads the whole repository is a second project. Give each role the minimum context it needs, not a shared copy of everything.
Put a budget on the group: maximum messages, maximum tool calls, maximum money. When the budget ends, stop with a partial result a person can read.
An unbounded committee of agents will talk until a rate limit talks for them. Measure cost per accepted result, including the human cleanup, or you will “save” money on a dashboard that ignores the afternoon someone spent repairing the run.
What to record before you rely on it
For a multi-agent design, record tokens and tool calls per role, not one blended number. Add the human minutes to repair ten runs.
Compare that total with a single-model workflow on the same ten inputs. Keep the design only if the accepted results justify the coordination.
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
Do more agents mean more points of view?
They mean more transcripts. Independent evidence comes from different data or a real check, not from a second copy of the same model with a different job title.
Should every step be its own agent?
No. A step with a fixed input and output is a function in a workflow. An agent is a cost you take on when the next action is not known yet.
Sources & further reading
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





