Releases·Tool Release
Subagent Dispatch Economics skill.md
A developer shared a SKILL.md, installable in .claude/skills/, that forces coding agents to run a fixed delegation test before spawning subagents. It was motivated by observed failures: agents spawning subagents for trivial greps, and under-briefed fresh agents re-deriving context the dispatcher already had. The skill covers a four-question delegation test, fork-vs-fresh shape selection, prompt scoping, wave sizing, and result verification.
9 engagement1 sourceTue, Sep 15, 2026, 08:56 AM
AI Overview
- Reddit user shares a SKILL.md that forces a four-question test (context cost, independence, reusability, judgment value) before an agent spawns a subagent
- Skill default favors forking over fresh agents, since forks inherit dispatcher context; fresh agents win only for unanchored second opinions
- Fresh-agent prompts must state goal and rationale, list ruled-out approaches, and draw explicit scope boundaries to avoid context re-derivation
- Waves are capped at three to five subagents, and completion claims must be verified against actual artifacts, not the subagent's report
- Tasks failing all delegation criteria can still batch into parallel dispatcher tool calls for wall-clock savings, but a lone trivial task stays inline
- Every dispatch decision, including not delegating, requires a one-line justification of shape and reasoning
Related
Social·Community
Blog post argues delegation in AI harnesses cuts token costs and enables distributed systems
A developer published a blog post reviewing AI subagents used in harnesses, arguing that scaling requires task delegation. Delegation saves tokens by limiting context per agent and enables distributed, non-monolithic systems. The post is based on the author's hands-on experience building such systems.
A developer published a blog post reviewing AI subagents used in harnesses, arguing that scaling requires task delegation. Delegation saves tokens by limiting context per agent and enables distributed, non-monolithic systems. The post is based on the author's hands-on experience building such systems.
7 engagement1 sourceredditTue, Jul 28, 2026, 03:31 PM
Social·Community
Redditor shares cost-cutting prompt that has Claude plan and delegate routine tasks to cheaper subagents
A Reddit user reports success with a prompt that reduces Claude Code usage costs by routing work between models: a top model handles planning and review while cheaper models execute routine sub-agent tasks. The approach targets cost per correctly completed task rather than raw output quality.
A Reddit user reports success with a prompt that reduces Claude Code usage costs by routing work between models: a top model handles planning and review while cheaper models execute routine sub-agent tasks. The approach targets cost per correctly completed task rather than raw output quality.
206 engagement2 sourcesredditSat, Sep 5, 2026, 12:26 PM
Products·building
Agent Prompt Architecture skill.md — skill for designing and reviewing agent system prompts
A Claude skill (skill.md) for designing and reviewing the prompts that run AI agents. It applies an architecture-level method treating system prompts as more than text blocks, targeting common agent failures: wrong tool selection, context drift on long tasks, and runs with no defined end condition. Aimed at people building agents.
A Claude skill (skill.md) for designing and reviewing the prompts that run AI agents. It applies an architecture-level method treating system prompts as more than text blocks, targeting common agent failures: wrong tool selection, context drift on long tasks, and runs with no defined end condition. Aimed at people building agents.
2 engagement1 sourceredditThu, Sep 3, 2026, 05:30 AM
Developments·Paper
SkillSV paper proposes structure-aware Shapley valuation for agent skills
A new arXiv paper introduces SkillSV, a structure-aware Shapley-style method for valuing the internal units of agent skills, such as rules, examples, scripts, and heuristics. It addresses the challenge of credit assignment in structured skill artifacts under automated feedback loops, which is relevant for optimizing agent performance.
A new arXiv paper introduces SkillSV, a structure-aware Shapley-style method for valuing the internal units of agent skills, such as rules, examples, scripts, and heuristics. It addresses the challenge of credit assignment in structured skill artifacts under automated feedback loops, which is relevant for optimizing agent performance.
0 engagement1 sourcearxivWed, Aug 5, 2026, 07:56 AM
Developments·Paper
Researchers introduce AgentLance for decentralized LLM agent orchestration
The paper introduces AgentLance, a decentralized orchestration method for LLM agents using repeated labor markets. It addresses bottlenecks, privacy issues, and manipulation in centralized systems. This is relevant for practitioners managing large-scale agent deployments.
The paper introduces AgentLance, a decentralized orchestration method for LLM agents using repeated labor markets. It addresses bottlenecks, privacy issues, and manipulation in centralized systems. This is relevant for practitioners managing large-scale agent deployments.
0 engagement1 sourcearxivMon, Aug 24, 2026, 10:15 PM