[P2] Experiment with depth >1 recursion and recursive agent harnesses #20

New issue

Closed

opened 2026-08-12 01:20:20 +00:00 by lost-rob0t · 0 comments

lost-rob0t commented

2026-08-12 01:20:20 +00:00

(Migrated from github.com)

Goal

Evaluate deeper recursion and recursive subagent composition only after the depth-1 runtime is measurable and stable.

Scope

enable depth >1 behind an explicit experimental flag;
compare nested RLM calls against delegated subagents and fresh-root artifact handoff;
enforce hard global budgets across the entire recursive tree;
test recursive capability narrowing and cancellation propagation;
measure quality/cost/latency against depth 0 and depth 1 baselines;
investigate Recursive Agent Harness patterns without assuming deeper recursion is always better.

Acceptance criteria

depth >1 cannot be enabled accidentally by default;
recursive trees terminate under global budgets;
child capabilities cannot exceed parent capabilities;
benchmark results show when deeper recursion helps, hurts, or is neutral;
a documented promotion rule defines whether any deeper-recursion strategy graduates from experimental status.

Dependencies

#11
#16
#17
#18
parent #3

Research

See RLM-RESEARCH-005-swi-agent-runtime.org, RLM-RESEARCH-008-adaptive-recursion.org, and the Recursive Agent Harness findings.

## Goal Evaluate deeper recursion and recursive subagent composition only after the depth-1 runtime is measurable and stable. ## Scope - enable depth >1 behind an explicit experimental flag; - compare nested RLM calls against delegated subagents and fresh-root artifact handoff; - enforce hard global budgets across the entire recursive tree; - test recursive capability narrowing and cancellation propagation; - measure quality/cost/latency against depth 0 and depth 1 baselines; - investigate Recursive Agent Harness patterns without assuming deeper recursion is always better. ## Acceptance criteria - depth >1 cannot be enabled accidentally by default; - recursive trees terminate under global budgets; - child capabilities cannot exceed parent capabilities; - benchmark results show when deeper recursion helps, hurts, or is neutral; - a documented promotion rule defines whether any deeper-recursion strategy graduates from experimental status. ## Dependencies - #11 - #16 - #17 - #18 - parent #3 ## Research See `RLM-RESEARCH-005-swi-agent-runtime.org`, `RLM-RESEARCH-008-adaptive-recursion.org`, and the Recursive Agent Harness findings.