01 /Tag — Ai Agents
20 articles in this topic.
An approval request should stop authorising new execution after its defined expiry. Resolve approval and timeout races in trusted state so a late click cannot unexpectedly activate stale work.
Several agents can advise on one business action, but one trusted owner should control its execution. Otherwise parallel recommendations can become duplicate writes or leave everyone assuming another agent completed the work.
A multi-agent conclusion should be traceable to its evidence, not merely to the last agent that repeated it. Preserve the lineage of consequential claims so reviewers can find unsupported upgrades and shared-source errors.
A human checkpoint should occur before the consequential effect, with enough evidence for a meaningful decision. Placing a person after several agents have already committed changes is review, not prior approval.
Conflicting agent answers need evidence-based resolution, not a vote on which response sounds more confident. First establish whether the agents are answering the same question about the same resource and policy version.
One unavailable agent should produce a defined workflow outcome, not an indefinite wait or a fabricated substitute answer. Decide which specialist outputs are required and which tasks can continue without them.
Agent memory should help interpret a task, while durable workflow state governs what may happen next.
Cost per completed multi-agent task should include unsuccessful runs, retries, specialist calls, and recovery effort. Counting only the final agent’s usage understates what the architecture spends to produce a correct result.
A caller correction must reach every agent whose work depends on the changed field. Updating only the conversational agent can leave a specialist returning stale options or an executor writing the superseded value.
Uncertainty can disappear when one agent summarises another’s output. Preserve whether a statement is observed, caller-reported, inferred, disputed, or unresolved instead of compressing every claim into a confident sentence.
Parallel agents can independently propose the same sensible action.
Multi-agent traces can duplicate personal information across prompts, handoffs, tool payloads, and summaries. Design review evidence around task relationships and safe outcomes instead of storing every message by default.
Route work to specialists through explicit task and permission boundaries. A specialist label does not prove expertise, grant access, or establish who may perform the final business action.
A recommendation from one agent is input to a decision, not permission for another agent to act. Preserve separate states for suggested, approved, executing, and completed work across every handoff.
Agents can share evidence by reference instead of copying entire conversations. Preserve source identity, access scope, and claim relevance so the receiving agent can verify what matters without accumulating unnecessary personal context.
Multi-agent budgets must apply to the shared task, including retries and delegation. Per-agent limits can still permit expensive loops when each handoff starts a fresh allowance.
An agent handoff should transfer a bounded task and its evidence, not merely a conversational summary. The receiving agent needs to know what is requested, what is already established, and which actions remain outside its authority.
A multi-agent workflow passes only when the final business state matches the authorised request. Agreement among agents or a convincing final explanation cannot prove that the correct customer record changed exactly as intended.
Agent handoff contracts need versions when field meaning or required structure changes. An older specialist can produce plausible output that the new coordinator interprets incorrectly unless the boundary is explicit.
One agent is often easier to operate when the task shares one context and one authority boundary. Add specialists only when they provide a measurable benefit that justifies more handoffs, latency, cost, and failure states.