Cognitive orchestration

An agent loop sounds simple until it runs. Read working state. Call a cognitive unit. Validate output. Update accepted meals. Maybe ask a human. Maybe patch one evening. Maybe stop because further iteration has little expected value. Memory reads interleave with deterministic gates. Authority checks sit between proposal and side effect. Logs must show who decided what, separate from what the model said. Vendor frameworks may help with adapters. The architecture itself is explicit design of who runs when, with what grants, toward which goal, with recorded transitions.

Helpful context: What is an agent owns the goal. What is a skill packages known procedures the orchestrator invokes. Libraries of cognition supply reusable cognitive units and Skills. The Thoughtware system situates orchestration among memory, behaviour, and evaluation.

One loop iteration, concretely

Follow the Meal Planning Agent through one material lap after CritiquePlan finds a weekday weakness. The agent reads working state: interpreted week, plan version, accepted meals, open gaps, loop budget. It acts by invoking RecommendMealSubstitution or the Compose Weekly Plan Skill depending on strategy. It observes the structured result, cost, source, and effect on state. Deterministic code then validates schema, allergies, and portions. The review step chooses accept, patch, fetch, ask, escalate, or stop. Finally the agent writes a transactional update to working state and checkpoints if the change is material.

Each step has an address in code and in evaluation, separate from chat history. Orchestration makes the loop inspectable: judgments, guards, memory, and handoffs toward a declared goal. Compare with a prompt that whispers "first check allergies, then suggest meals, then update memory." The whisper may work in a demo. It fails incident review because nobody can point to the transition that authorized a purchase or skipped a guard. The difference between these approaches becomes vivid during the first production incident: with explicit orchestration, operators identify that the allergy guard fired after the patch step rather than before it, and the fix is a reordering in code rather than a prompt rewrite that may introduce new regressions elsewhere.

What orchestration coordinates

Orchestration sits above individual calls. It answers sequencing questions the calls cannot answer alone. cognitive units supply bounded judgments. Skills supply known procedures. Deterministic code enforces exact rules and executes permitted side effects. Memory separates context, knowledge, experience, expertise, and working state. Human handoffs occur through targeted questions when consent, preference, or authority is missing.

Libraries make composition safe. A domain library exposes GenerateCandidates and AssessMealPracticality. A judge library exposes JudgeWeekdayPracticality. Agent-private inventory holds strategy-specific cognitive units like InterpretWeek and CritiquePlan. Orchestration chooses among packages with search and substitution rules, not ad hoc prompt paste.

Orchestration concern	Wrong home	Right home
Allergy enforcement	Inside a generative call	Deterministic shell on endorsed knowledge
When to stop iterating	Token limit	Agent stop policy tied to goal
Which procedure to run	Model improvisation	Skill invocation with contract
Missing household preference	Guessing	AskTargetedQuestion with bridge transition

What orchestration is not

Cognitive orchestration is not a longer system prompt ordering tools. Prompts may describe behaviour. They do not replace logged transitions. It is not a DAG label without stop logic, because pipelines are not agents and fixed graphs may live inside Skills as perform steps. It is not composite cognitive units smuggling several judgments into one call, because each judgment deserves its own contract and suite. And it is not autonomy without libraries, because searchable packages make composition safe at scale. See libraries of cognition.

Orchestration connects to trust conduct: stop conditions as design, targeted questions, working state versus transcript. Trust failures often trace to missing orchestration, not to weak phrasing inside one call. The same pattern repeats across domains. Enterprise intake runs the same choreography with different nouns: read case state, invoke extraction Skill, run exact validators, call field cognitive units for ambiguous interpretations, review whether to patch, escalate, or close, write state transactionally. The domain changes. The choreography grammar does not.

Orchestration makes the loop inspectable: judgments, guards, memory, and handoffs toward a declared goal.

Thoughtware White Paper S06

Ordinary code, explicit transitions

An agent is mostly ordinary code because orchestration must survive staff turnover and vendor refresh. Engineers should read the loop in an afternoon. Operators should see which transition fired after a failed guard. Explicit review verbs give the loop decision locality. Evaluators can ask whether the agent repaired locally, fetched missing knowledge, replanned, or continued without reason. That evidence requires orchestration, not fluency.

Common failures

Hidden ordering is the most frequent failure: steps exist but only in prompt prose, and incidents become archaeology. Memory discipline collapse follows, where temporary corrections write directly into endorsed knowledge without approval. Guard skipping after a fluent proposal occurs when code paths exist but triggers attach to the wrong stage. Human handoffs as generic chat means missing preference becomes small talk instead of a targeted question that changes the plan. Each failure has an orchestration fix: move the rule to code, name the memory form, attach the guard to the correct stage, use AskTargetedQuestion with a bridge transition.

Observability expectations

Operators should replay one iteration from logs without opening prompt files. That standard forces orchestration discipline. Each step emits which dependency ran, which guard passed or failed, which review transition fired, and which working state fields changed. Dashboards that show only token spend hide orchestration health. Pair cost metrics with transition counts: patches versus replans, asks versus guesses, stops versus cap hits. Spikes in replan often indicate representation problems upstream, not model weakness downstream.

Human participation belongs in orchestration by design. When AskTargetedQuestion fires, the loop records what gap the question fills and what transition occurs when the answer arrives. Bridge transitions exist so pursuit continues without turning the household into a permanent chat partner. The same discipline applies to escalation. An escalated case leaves a trace in working state: why leadership returned, what authority was exceeded, what remains blocked until a human acts.

Memory reads in sequence

Orchestration must respect memory forms. Context supplies this run's request. Knowledge supplies endorsed facts like allergies. Working state supplies pursuit progress. Orchestration that reads knowledge as context risks treating durable facts as disposable. Orchestration that writes context into knowledge without approval risks silent policy drift. Sequence memory operations explicitly in the iteration sketch so reviewers see which form each read or write touches.

When a lap fails, logs should isolate whether the cognitive unit, guard, memory write, or review transition failed. Orchestration without isolation sends every incident to "the model behaved oddly." Isolated logs shorten repair. Build failure isolation into the iteration template before scale, not after the first production incident. Teams that instrument each orchestration step from the beginning find that most production defects resolve in minutes because the log already names the responsible component, while teams without isolation spend hours reproducing failures that were never recorded at the correct granularity.

What to do next

The orchestration discipline becomes concrete when one iteration is drawn with numbered steps and named owners. Mark every side effect with the grant that permits it. List memory reads and writes by form: context, knowledge, working state. Replace prompt ordering with logged transitions reviewers can replay. Confirm the lap template includes working state read, dependency call with contract id, guard execution log, review transition enum, transactional state write, and checkpoint when material. Missing any item means orchestration is still aspirational in comments rather than enforced in code.

See where the cognitive unit ends and the agent begins, the agent loop, and minimum sufficient loop.

Read next: Where the cognitive unit ends and the agent begins.