Three execution forms, not a maturity ladder
Progressive Crystallization · Part 2 of 10
A playbook can be reusable and still ask an agent to choose every next step. Calling it a playbook tells us that someone captured a way of working. It does not tell us who controls execution.
That distinction is easy to lose when discussing automation as a progression. We draw a line from exploration to rules, put the rules at the far end, and quietly turn an architectural choice into a score. Then a team with a useful hybrid workflow looks unfinished.
I would rather ask two questions: who chooses the next step, and which steps still invoke a model? They give us three useful execution forms. The names matter more than the numbers.
The same question, three execution forms
Consider a fictional lab. A responsible operator wants to know whether a permitted interface appears up or down. The example is independently constructed for this series. It is not a production incident, and there is no authority to change the interface.
In an agent-led design, an agent chooses how to investigate within its permitted scope. It might inspect a status response, notice that its meaning is unclear, and ask for an approved explanatory document. The agent helps the directly responsible individual, or DRI, interpret the evidence. It can follow a reusable playbook while still deciding which permitted observation to request next. A captured sequence does not make the orchestration deterministic.
In a hybrid design, code fixes the procedure: check eligibility, collect a status response, ask a model to interpret it, validate the answer, then follow an explicit branch. The model might return “appears down” with an evidence reference. That answer selects a reporting branch. It cannot invent a reset operation. The reasoning is bounded, but it still affects what happens.
In a deterministic design, the eligible response already has an agreed status field. A parser checks the contract and maps the known values to a report, without runtime LLM calls anywhere inside that stated workflow. An unsupported value stops the path. There is no benefit in adding a model to rediscover an enum the operator has already specified.
These descriptions classify execution, not correctness. A deterministic parser can misread the field. An agent can reach a correct conclusion. Neither fact changes the form.
Observed responses, competing explanations, unresolved questions.
A domain expert's scenario, procedure, and documented assumptions.
Either source informs a reviewed contract and existing authority.
Define the task, eligible evidence, allowed outcomes, owner, and conditions for stopping. Select any suitable form below.
Parallel choices, not transitions between stages.
A model directs permitted investigation and tool use.
A fixed procedure contains bounded model decisions.
Explicit rules execute without runtime AI in the declared scope.
Location is a separate decision: retain a playbook, or incorporate understood behavior into service code where appropriate.
Stop the affected path. Reconsider its contract or form under separately granted authority, without an automatic privilege increase.
Use names first, numbers second
Section III of my Progressive Crystallization paper, version 1 names agent-orchestrated, hybrid, and deterministic playbooks. Its numbering runs from Type 3 for agent orchestration to Type 1 for deterministic execution. For readers using V1, V2, and V3 as shorthand for a possible development story, the order is reversed.
- Agent-led
- Paper Type 3. Sometimes called V1 in developmental shorthand.
- Hybrid
- Paper Type 2. Sometimes called V2.
- Deterministic
- Paper Type 1. Sometimes called V3.
I use the form names throughout this series to avoid making that shorthand carry an argument it cannot support. “Type 1” is not a rating of operational maturity. “V3” is not a promise that a particular system should arrive there.
The framing here also makes a deliberate distinction from a literal reading of the paper's lifecycle: investigation is one source of knowledge, not a compulsory first stage. A well-understood procedure can start deterministic. A workflow with durable interpretive needs can remain hybrid. The taxonomy is useful without assuming automatic promotion or a universal safety improvement.
Anthropic's Building effective agents provides a related distinction. It calls systems with predefined code paths workflows, while agents let models direct their process and tools. Those definitions help describe orchestration. They do not, by themselves, distinguish a fixed workflow with a model call from one with no runtime inference. That is why the hybrid category is useful here.
Draw the boundary before choosing a label
Suppose an agent-led investigation calls a reusable status checker. The checker always collects the same permitted fields and asks a model to interpret an operator note. Its outer procedure is fixed, so the checker is hybrid. The collector inside it is deterministic. The overall investigation is agent-led.
The agent chooses whether to invoke the status checker.
Fixed collection, interpretation, validation, and reporting sequence.
Fixed permitted observation.
Interpret the supplied note or abstain.
Write the boundary in the design record. Include retries and recovery. A checker that normally uses rules but silently calls a model on an unfamiliar value is not entirely deterministic. A checker that stops and opens a separately authorized investigation can be deterministic within its own scope, provided the wider process is described honestly.
Human review is a separate dimension too. A deterministic operation may require approval before it runs. An agent-led investigation may be permitted to read a narrowly approved fixture without repeated prompts. Neither example establishes a blanket rule for reads or writes. Reading can expose confidential data or impose load; a one-word classification can trigger a consequential action if the surrounding design permits it.
Choose for the uncertainty that remains
For the fictional lab's documented enum, start with a parser. For a stable procedure containing an ambiguous operator note, consider hybrid execution. For a new diagnostic question where the next useful observation depends on the last one, agent-led investigation may be appropriate, subject to scope and authority.
And these can coexist. The routine status question can use rules while a different, authorized investigation explores why the interface is down. Reporting status is not diagnosing cause, and diagnosing cause is not permission to mitigate it.
A useful review asks whether the selected form fits the task's remaining uncertainty. It does not ask why every box has not yet moved to the right.
Sources and scope
All traces above are hypothetical. No model comparison or production evaluation was performed. This article offers design distinctions, not a novelty claim or an empirical ranking of forms.
- Arun Malik, Progressive Crystallization, arXiv v1, Section III: source for the three execution-type names and paper numbering. The series uses a nonmandatory selection model and does not reproduce the paper's operational measurements or validate its guarantees.
- Anthropic, Building effective agents: “What are agents?”, “When (and when not) to use agents”, and “Combining and customizing these patterns.” Vendor architectural guidance, not evidence that the fictional designs are equally effective.