Bound delegation depth and tool-call volume

Recursive delegation can grow faster than expected: each worker creates more workers, which repeat searches and consume budget. Make depth and volume first-class state.

Check a bounded tree

The agent controls lab creates an authored five-node tree.

from agent_cases import delegation_case

result = delegation_case()
assert result["observed"] == {"depth": 2, "nodes": 5, "tool_calls": 5}
assert result["within_bounds"] is True
assert result["further_delegation_allowed"] is False

Limits are depth 2, six nodes and eight tool calls. The simulation does not spawn agents.

Allocate child budgets

The parent grants each child a subset of remaining time, calls and cost; children cannot create budget. Attach task ID, scope, allowed tools and expected output. Reject duplicate or cyclic delegation.

Use a global coordinator for counters and cancellation. Stop new work as limits approach and preserve partial results. Parallelism does not remove the end-to-end deadline.

Evaluate usefulness

Count useful evidence and accepted outputs, not only completed branches. Penalize duplicate searches and handoff loss. Compare against bounded single-agent and workflow baselines.

The Generative & Agentic AI course links delegation bounds to trajectory evaluation and partial-completion runbooks.

Exercise

Implement a coordinator with depth, node, call and time caps. Attempt cycles and budget amplification; prove every branch terminates and the final report lists incomplete work.

Continue learning

This article is part of the Agent workflows and state sequence. Use the neighbouring tasks when you need the prerequisite or the next application.

• Review the prerequisite or neighbouring task in Test a multi-agent handoff against a single-agent baseline.
• Continue with Evaluate an agent trajectory as well as its final answer.

Reference: NIST Generative AI Profile.