Managing multiple AI agents means deciding which jobs can run independently, which must wait and who may change a shared resource. One coordinating conversation can help, but the coordination rules determine whether the work remains understandable as the number of workers grows.
The original Growth Cab post described managing many conversations through one. The useful lesson is the management problem it exposed: a worker can finish its assignment while the overall result remains incomplete. This guide focuses on concurrent work, dependencies and recovery around those handoffs.
in
“I have +50 Claude chats running every day. I manage all of them through one.”
Separate Independent Work From Dependent Work
Map the inputs and outputs before dispatching tasks. Account research and analysis of yesterday's campaign replies may run independently. Drafting messages from that new research depends on the research being accepted first. Starting both together can cause the writer to use old data or invent the missing input.
Write the dependency in the task record. A waiting task should identify the exact artifact and version it needs. When the source changes, decide whether the dependent output needs another review. A completion message from one worker should never silently authorize another to act on an unreviewed result.
Give Every Task One Owner
Record a task identifier, expected output, worker, reviewer, input version and destination. Use distinguishable states such as queued, running, waiting, needs review and completed. Name the evidence required for completion. A sentence saying the work is done is a status claim until the destination or artifact has been checked.
Assign one owner to reconcile the overall result. The coordinator should know which outputs were accepted, which were rejected and which remain uncertain. This role can organize review through software, but consequential decisions still need the authority specified for the task. Coordination does not enlarge a worker's permissions.
Control Changes to Shared Resources
Two agents can safely read the same approved reference while producing separate drafts. Two agents editing the same CRM record or publishing the same page require explicit coordination. Assign a single writer or use isolated copies with a defined review and integration step. Preserve the original state so competing changes remain recoverable.
Consider a fictional case with one agent researching an account and another correcting its CRM record. Both may discover a new company name. The coordinator must reconcile identity and source evidence before either change becomes authoritative. Letting the last write win would hide the disagreement rather than resolve it.
Review Handoffs Against the Whole Goal
Check that each artifact answers its assigned question and fits the next step. A research document may be accurate but omit the field the writer needs. A draft may be readable but use an old offer. Review both local correctness and the compatibility of connected outputs.
Keep a compact record of accepted artifacts, source versions, checks and unresolved gaps. This allows a replacement coordinator to recover the work without rereading every conversation. Reuse good briefs deliberately after review; memory does not improve itself merely because several agents share a manager.
Recover Stalled and Uncertain Tasks
A timeout does not prove that an action failed. Inspect the destination before retrying a write or external action. Otherwise a recovery attempt can duplicate work that already succeeded. Preserve an uncertain state until the check establishes what actually happened.
When a worker loses context, pass a checkpoint containing the goal, permissions, completed work, evidence and next unfinished step. Do not restart accepted work automatically. When a dependency fails, mark downstream jobs as waiting and explain what would unblock them. Quietly skipping them creates a false completion report.
Measure Coordination Overhead
Track accepted outputs, rework caused by handoffs, conflicts, waiting time and review effort. Adding another worker is useful when it shortens the overall path without increasing unresolved errors. If most jobs depend on the same unfinished input, additional parallelism may add little value.
Finish with one report that identifies the final artifact, the checks performed and any remaining gap. Distinguish individual task completion from the completed business outcome. A set of successful worker messages is insufficient when the assembled result has never been inspected.
Use the business operating layer guide to design the broader interface, context and permissions.
Use the task delegation guide to prepare an individual brief.
Use workflow regression checks when changes may affect previously accepted behavior.
