Hr onboarding · Production

Designing decision-heavy enterprise AI systems without losing control: tri-agent architecture (Planner–Orchestrator–Executor)

The problem

Enterprise AI workflows involving complex, multi-step decisions cannot be handled by single-agent architectures, which failed to enforce execution order, track state, or prevent repeated and skipped steps under real-world conditions.

First attempt

Three successive designs failed: (1) a monolithic agent executed steps out of order and lacked state tracking; (2) upgrading to a more capable model increased latency and made the agent overly cautious; (3) delegating planning to a separate component still allowed parallel execution that violated sequential dependencies.

Workflow diagram · grounded in source
1
User request triggers workflow
trigger
“A manager might request: "Onboard a new employee, Jane Doe, as a Sales Representative."”
2
Planner produces validated plan
ai_action
“A reasoning agent that interprets the user's request and produces a validated plan. The plan is a structured list of actions, complete with their order, pre-checked prerequisites, required parameters, and placeholders for any missing inf…”
3
Orchestrator loads plan into task queue
integration
“The Orchestrator loads the plan into the queue, steps through tasks one by one, and can pause the process if needed”
4
Orchestrator dispatches and controls execution
routing
“It decides when to invoke the Planner (e.g., on a new request or after new information comes in), when to dispatch the next task to the Executor, and how to react to the Executor's results”
5
Executor carries out task
ai_action
“A focused agent that carries out one task at a time (or multiple independent tasks in parallel). It takes the next task from the queue (already fully specified by the plan, including which tool/API to use and with what parameters) and ex…”
6
Result returned for evaluation or re-planning
feedback_loop
“After execution, it returns a structured result (or error) back to the Orchestrator”
Reported outcome

The tri-agent architecture (Planner–Orchestrator–Executor with a task queue backbone) delivered deterministic, ordered execution of decision-heavy enterprise workflows; pruning the Executor's toolset cut execution latency roughly in half.

Reported metrics
Execution latency from tool pruningroughly in half
Reported stack
LLMtask queue
Source
https://medium.com/data-science-at-microsoft/designing-decision-heavy-enterprise-ai-systems-without-losing-control-60ac5d25e1e3
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

The tri-agent architecture (Planner–Orchestrator–Executor with a task queue backbone) delivered deterministic, ordered execution of decision-heavy enterprise workflows; pruning the Executor's toolset cut execution lat…

What tools did this team use?

LLM, task queue.

What results were reported?

Execution latency from tool pruning: roughly in half (source-reported, not independently verified).

What failed first in this deployment?

Three successive designs failed: (1) a monolithic agent executed steps out of order and lacked state tracking; (2) upgrading to a more capable model increased latency and made the agent overly cautious; (3) delegating…

How is this hr onboarding AI workflow structured?

User request triggers workflow → Planner produces validated plan → Orchestrator loads plan into task queue → Orchestrator dispatches and controls execution → Executor carries out task → Result returned for evaluation or re-planning.

Related hr onboarding cases