agent-graph-designer

Agent Graph Designer

Design the graph around guarded loops. Treat nodes as bounded work contracts and edges as evidence-backed product decisions.

Step 0 — Preserve product judgment

Before choosing a topology, ask only for missing parts of the user's own:

1. problem framing;
2. hypothesis and proposed answer;
3. outcome North Star, leading metrics, guardrails, and key trade-off;
4. proposed next step.

Critique contradictions after the user answers. Never replace these decisions with a fashionable topology.

Step 1 — Qualify the shape

Read references/qualification.md. Return exactly one verdict before drawing anything:

• LOOP_SUFFICIENT: one objective, one primary working context, sequential execution, and no meaningful fan-out/fan-in or conditional ownership boundary. Route to loop-designer when it is recurring.
• GRAPH_REQUIRED: at least one structural graph need and one coordination need are evidenced.

Structural graph needs:

• two or more independent branches can run before a deterministic join;
• evidence can route the workflow to materially different states;
• multiple bounded loops must coordinate without sharing one unbounded context.

Coordination needs:

• distinct owners, tools, permissions, budgets, or verifier independence;
• typed artifact handoffs or shared state are required;
• failure in one branch must block, degrade, compensate, or escalate differently.

Do not invent parallelism. A sequential checklist remains a loop or pipeline.

Step 2 — Lock the graph outcome contract

Emit:

GRAPH DECISION
Verdict: LOOP_SUFFICIENT | GRAPH_REQUIRED
Problem:
Hypothesis:
Outcome North Star:
Leading metrics:
Guardrails:
Key trade-off:
Proposed next step:
Qualification evidence:

The North Star must measure a real outcome, not graph creation, node count, agent count, or usage.

If the verdict is LOOP_SUFFICIENT, stop after explaining the smaller design and route to loop-designer. Do not emit a decorative graph.

Step 3 — Define node contracts