Two agentic skills: find the highest-leverage move, then plan the wave
An optimizer that finds the single highest-leverage move in any context, and an orchestrator that plans a mission into a wave-based DAG of parallel sub-agents, gated by an auditor.
Role separation and autonomy tiers done deliberately, across 7 agents, standalone and public. Two skills that show the same discipline from opposite ends: one narrows a whole situation to a single move, the other fans a mission out across parallel agents and gates the result.
The problem
Two recurring failure modes show up when you point agents at real work. The first is doing too much at once: a long, undirected effort that never names the one move that actually matters. The second is the opposite, a single agent trying to carry a whole multi-step mission alone, with no separation of roles and no checkpoint, so it drifts or loops with nothing to stop it. These two skills are deliberate answers to each: one forces focus down to a single highest-leverage move, the other spreads the work across parallel sub-agents and puts an auditor in the path.
Approach
The two skills are built as standalone tools with explicit role separation and autonomy tiers, not one monolith. The optimizer reads a whole context and returns the single move with the most leverage, rather than a long list. The orchestrator plans a mission into a wave-based DAG, runs sub-agents in parallel where it is safe to, and routes the output through an auditor before a wave is allowed to count as done. Capability is the easy half; the design work is the gates and the stopping rules.
- Find one move, not a backlog. The optimizer reads the full context and names the single highest-leverage action. The leash: it surfaces the move for a human to take, rather than firing it off itself.
- Plan in waves, run in parallel. The orchestrator decomposes a mission into a DAG and runs independent sub-agents at the same time. The leash: each wave is gated by an auditor before the next one is allowed to start.
- Separate the roles, tier the autonomy. Each agent has a defined role and a defined level of independence. The leash: a hard safety floor and a run budget bound the whole thing, so it cannot quietly run away.
What was built
Two standalone skills, spanning 7 agents in total. The optimizer is a single-purpose skill that collapses a situation to its highest-leverage move. The orchestrator is a wave-based planner that turns a mission into a DAG of parallel sub-agents with an auditor in the path. The role separation, the autonomy tiers, and the auditor gate are the deliberate parts: they are what keep a parallel, multi-agent run honest and bounded instead of sprawling.
Guardrails
What the system is structurally not allowed to do. This is the through-line: capability, then leash.
- NoIt cannot let a wave count as done unaudited. An auditor gates each wave before the next begins.
- NoIt cannot run without limit. A run budget stops a runaway loop before it spirals.
- NoIt cannot drop below the hard safety floor. That floor binds every agent regardless of tier.
Stack and tools
My role
I designed and built both skills, the optimizer and the wave-based orchestrator, including the role separation, the autonomy tiers, and the auditor gate that bounds each run.
Links and verification
This is an open-source project. The repository is public, so the architecture and guardrails described here can be inspected directly.