Skip to main content

Toolkit

Agents & Subagents · Pitfalls

On this page7 sections

Pitfalls

Delegating an Unbounded Goal

“Improve the codebase” creates scope drift and an unreviewable result. Assign one objective and a stop condition.

Concurrent Writes Without Ownership

Context isolation does not prevent file conflicts. Allocate files or isolate worktrees.

Authority Leakage

A specialized role may inherit tools or credentials unrelated to its task. Apply least privilege at the runtime layer.

Context Dumping

Returning every log and file defeats the point of delegation. Require a structured result with links to evidence.

Blindly Trusting Completion

The parent must inspect the diff and rerun the decisive checks. Child confidence is not integration evidence.

Deep Nesting Without Observability

Nested delegation can obscure who acted, where tokens went, and which instruction set applied. Bound depth and preserve causal identifiers.

Falsification Checklist

  • Could the task have remained in the parent with less overhead?
  • Is write ownership disjoint and explicit?
  • Can the child act outside the user's authorization?
  • Does the result contain enough evidence to verify independently?
  • Can every nested action be attributed to a run and role?

Source register

Agents & Subagents sources

Primary vendor documentation reviewed on the date shown. Links may change after publication.

  1. Create custom subagents (opens in a new tab)

    Anthropic · official docs · accessed 2026-08-01

  2. Subagents (opens in a new tab)

    OpenAI · official docs · accessed 2026-08-01

  3. Overview of customizing GitHub Copilot CLI (opens in a new tab)

    GitHub · official docs · accessed 2026-08-01

  4. GitHub Copilot CLI command reference (opens in a new tab)

    GitHub · official docs · accessed 2026-08-01