The coordinator is the most strategically important component in an AI production system. It decides when a run begins, which state is accepted, what may happen next, when a model or tool is invoked, whether approval is required and when the process is complete.
That role can be delegated to an LLM. It can also remain in deterministic code or be distributed across heuristics, state machines and model calls. The choice determines how much of the production process the organization can inspect, govern and replace.
Coordination Is More Than Delegation
An agent manager that assigns work to specialist agents is one form of coordinator. Production coordination is broader. It owns the run lifecycle:
load accepted state
→ determine available transitions
→ invoke code, model or human
→ validate the proposed result
→ execute a permitted side effect
→ record the accepted transition
→ continue, pause, compensate or finish
The coordinator does not need to perform every task. Its value comes from deciding which task becomes part of the governed process.
This creates the central architecture question:
Does the model coordinate the production system, or does the production system coordinate bounded model operations?
Three Coordinator Architectures
Model-directed coordinator
An LLM receives the goal and current context, decomposes the task, selects tools or agents, evaluates intermediate results and decides when to stop. This works well for ambiguous research, exploration and other tasks where the useful path cannot be enumerated economically.
The tradeoff is that routing, delegation and completion criteria become probabilistic. A trace can show which tools were called, but the governing process exists largely as the behavior of one model run. Model updates can change not only output quality but the process itself.
Heuristic coordinator
A heuristic coordinator uses rules, scores, thresholds, queues or search policies. It can route a low-confidence result to review, select a specialist based on document type or stop when marginal improvement falls below a threshold.
Heuristics support variation while keeping the decision mechanism inspectable. They become difficult when many rules interact, but they remain testable against known cases and can explain which threshold or policy selected a path.
Deterministic coordinator
A state machine, execution graph, actor system or coded workflow defines legal states and transitions. Given the same accepted state and event, it selects the same transition. This is appropriate for approvals, financial actions, publishing, entitlement changes and other processes where order and accountability matter.
Deterministic coordination cannot encode every open-ended judgment without becoming brittle. Its purpose is not to eliminate models, but to bound where probabilistic choice enters the process.
The Hybrid Pattern
Most production systems need all three mechanisms:
deterministic shell
├── heuristic routing for known variability
├── model operators for interpretation and generation
├── explicit approval and policy gates
└── authoritative state transitions
The deterministic shell defines identity, legal states, budgets, timeouts, tool boundaries, completion conditions and compensation behavior. Heuristics choose among known branches. Models classify evidence, generate candidates or propose a plan where explicit rules would be uneconomical.
The model may recommend the next action. The coordinator determines whether the action is permitted, adequately supported and still within the run’s constraints.
This is why agentic workflows and production pipelines should not be treated as mutually exclusive. An agent can be a probabilistic operator inside a durable process.
Traceability Requires an External Acceptance Boundary
Tool-call logs provide observability, but not necessarily control. If the same model chooses the task, evaluates its own result and decides that the process is complete, the trace records what happened without providing an independent definition of what should have happened.
An external coordinator can record a transition decision:
accepted current state
→ transitions permitted by policy
→ evidence and model proposal
→ heuristic, rule or approval applied
→ selected action and tool result
→ validated new state
This preserves three different capabilities:
- ex ante control: inspect and test allowed transitions before a run;
- runtime control: pause, reject, retry or compensate a concrete step; and
- ex post traceability: reconstruct why authoritative state changed.
The coordinator should record observable reasons and evidence, not private model reasoning. Provenance, inputs, proposed actions, policy outcomes and accepted transitions are sufficient for operational explanation.
Akka as a Coordination Substrate
The actor model is a useful architectural contrast. An actor owns explicit state, receives messages and exposes defined behavior. Supervision makes failure handling part of the system rather than another instruction in a prompt.
Akka is therefore not simply an alternative to agentic AI. It can surround model operators with deterministic coordination. An actor may own a run or domain entity, call an LLM for an uncertain classification, validate the returned candidate and emit only an accepted event. The comparison of Akka and agentic workflows develops this distributed-systems perspective.
Durable workflow engines, execution-graph runtimes and event-sourced state machines can provide the same strategic property: process state and transition authority remain external to the model.
Process Sovereignty Is a Strategic Asset
Process sovereignty means that the organization owns the executable definition of how work becomes accepted. It does not merely own the prompt or a transcript of model behavior.
External coordination creates strategic options:
- models and providers can be replaced without redefining the business process;
- transitions remain testable when model behavior changes;
- run histories remain comparable across model versions;
- tools, skills and policies can evolve independently;
- approval and compliance evidence remains system-owned; and
- the organization retains the ability to intervene, replay and migrate.
This is the production-architecture consequence of The Production Black Box . A model’s internal generation process can remain opaque while its assignment stays bounded. The strategic risk rises when the model also absorbs coordination, acceptance and state transitions, because the smallest controllable production unit grows with it.
A Coordinator Decision Checklist
Keep a decision outside the model when it changes authoritative state, triggers an irreversible side effect, requires independent approval, must be replayable or determines whether the process is legally complete.
Delegate to a model when the decision depends on ambiguous language, open-ended evidence or creative synthesis—and when the returned result can still be validated before acceptance.
The objective is not maximum determinism. It is maximum useful model capability without giving up the ability to understand, control and replace the process that contains it.