execution-model

name	execution-model
description	Decides whether an automation should be built as a deterministic Make.com scenario, as an AI Agent, or as a hybrid (scenario shell with an AI Agent at the one indeterministic step). Produces a Routing Verdict recorded in the plan. Called before any blueprint is designed.
allowed-tools	Read, Write, AskUserQuestion

Skill: execution-model

Every automation gets one routing decision before any blueprint exists. Making it explicitly — and writing down why — is what stops the two expensive mistakes: an LLM call on a path that a filter could have decided, or a rigid router on a path that needs judgement.

The test

Apply per automation, not per project:

Can every branch be enumerated at design time as a filter or router condition over known fields?

• Yes → scenario. Deterministic. Reproducible, auditable, 1 operation per module, no model cost, no drift.
• No → the indeterministic part needs an AI Agent. A branch requires interpreting unstructured input, or the sequence of tools to call is not knowable ahead of time.

Then ask the second question, which decides between ai-agent and hybrid:

How much of the flow is actually indeterministic?

• One step (classify, extract, summarise, draft) → hybrid
• The whole control flow — the agent decides which tools to call, in what order, and when it is done → ai-agent

hybrid is the expected default

Most real automations are a deterministic pipeline with exactly one judgement call in the middle: a deterministic trigger, an AI Agent module at the one step that needs interpretation, then deterministic routing on its structured output.

Wrapping an entire deterministic pipeline in an agent is the common failure. It costs a model call per operation, is non-reproducible run to run, and turns a filter bug — which you can read — into a prompt bug, which you cannot. When in doubt, shrink the agent's job until only the genuinely indeterministic step remains.

Corollary: an AI Agent step should return structured output so everything downstream of it is deterministic again.

Signals

Points to scenario	Points to an agent
Trigger fields are structured (webhook payload, row, record)	Input is free text, email body, transcript, image, PDF
Branches map to known enum values or thresholds	"Decide which…", "figure out whether…", "route it appropriately"
Same input must always produce the same output	Output is a judgement, a summary, or drafted prose
Auditability or compliance matters on this path	The number of steps varies per input
High volume — per-run cost dominates	Low volume, high per-item value

Ambiguity is not a tie-break in favour of the agent. If you cannot name the specific step that needs judgement, the verdict is scenario.

Output — the Routing Verdict

Emit this block verbatim into the plan, and into .make/context/ai-agents.md when the verdict is not scenario:

Routing Verdict: scenario | ai-agent | hybrid
Indeterministic step: {the one step, or "none"}
Reason: {one line — what could not be enumerated at design time}

Set ai_required from the verdict: true for ai-agent and hybrid, false for scenario. This replaces keyword matching — a plan that merely mentions "summary" in prose is not an AI requirement.

Downstream contract

• plan-builder — carries routing_verdict on the AutomationPlan.
• ai-agent-designer / ai-agent-builder — run only when the verdict is ai-agent or hybrid. On a scenario verdict, refuse and cite the Reason line; if the user wants an agent anyway, re-run this skill with their new information rather than overriding the verdict silently.
• blueprint-review — on hybrid, confirm the AI Agent module returns structured output and that the downstream routing reads that structure rather than raw text.

Worked examples

Ask	Verdict	Why
New Stripe charge → append a row in Sheets	scenario	Every field is known and mapped; no branch needs judgement
Inbound email → route to the right team	hybrid	Trigger and delivery are deterministic; classifying the body is not
Watch a shared inbox, research each request, use whichever tools it needs, reply	ai-agent	Neither the tool sequence nor the step count is knowable at design time
Form submission → if amount > 5000 route to approver, else auto-approve	scenario	A threshold is a filter, not a judgement