Skip to main content

Anti-patterns

These anti-patterns show up repeatedly in early-stage agent systems. COP and ARP exist largely to avoid them.

“Mega-planner-agent with every tool”​

Putting every possible tool in the model’s action space:

  • makes selection unreliable,
  • makes evaluation hard,
  • increases blast radius and security risk.

ARP’s answer: bounded candidate sets per step.

“Prompt-as-policy”​

Relying on a system prompt to enforce safety or governance is fragile.

ARP’s answer: explicit policy checkpoints enforced by the Run Coordinator (and optionally delegated to a PDP).

“No durable evidence”​

If you can’t answer “what happened and why?” you can’t operate the system.

ARP’s answer: durable RunEvents (NDJSON) and ArtifactRefs (refs + blobs).

“One model does everything”​

When one LLM is responsible for planning, tool choice, and argument generation, failures become opaque and hard to debug.

JARVIS splits this into separate steps/services:

  • planning (Composite Executor),
  • candidate generation (Selection),
  • binding and argument generation (Composite Executor).