How Agents Run

Your agent is defined as static YAML, but it does not run like a static script. At runtime it reasons, calls tools, and routes work — no graph to draw.

Static definition, dynamic runtime#

A Prix agent is declared in YAML: its models, tools, context, and prompt. That part is static — there is no workflow graph, no node editor, no if X then Y wiring to maintain. What makes an agent “dynamic” is the runtime: the model decides what to do next on each turn.

	Workflow tools	Prix agents
Definition	A graph of nodes and edges	Static YAML + a prompt
Control flow	You wire every branch	The model chooses each step at runtime
New situation	Add a node or a rule	Reasoned through with the existing prompt
Specialist work	Another node in the graph	Delegated to a sub-agent on demand

The agentic loop#

Execution is a loop, not a pipeline: the model receives the prompt and context, decides on an action (usually a tool call), the runtime executes it, the result is fed back, and the loop repeats until the task is done or the turn budget is reached. The cap is max_turns in agent.yaml.

prompt + context
   │
   ▼
 model  ──▶  decides an action (tool call)
   ▲                 │
   │                 ▼
 result  ◀──  runtime executes the tool
   │
   └─ repeat until done, or max_turns reached

Because the path is chosen per input, an agent handles variance a fixed workflow cannot — it does not break when the world changes, it reasons about the new case.

Delegation: agents that route to specialists#

A single agent does not have to do everything. With the delegate tool, a parent agent can hand a sub-task to a specialist sub-agent at runtime. Routing is driven by each sub-agent’s description, not a hardcoded edge — the parent picks the right specialist for the input in front of it.

agent.yaml

tools:
  - name: delegate
    config:
      max_depth: 2

delegate_agents:
  summarizer:
    prompt: summarizer-prompt.yaml
    tools:
      - web_fetch
      - artifacts

This is the honest meaning of “dynamic workflow” on Prix: a reasoning loop plus runtime delegation. There is no visual graph to keep in sync with reality.

Multi-Agent

Compose a parent agent with specialist sub-agents.

Prompt Design

Write prompts that let the agent reason, not just follow steps.