Agent Workflow Controller

Design and audit complex multi-agent workflows with rigorous ownership, evidence gates, and failure recovery policies.

Install
cmdop skills install agensi-agent-workflow-controller

Agent Workflow Controller is a skill built for developers and AI orchestrators who need to manage complex, multi-step, or multi-agent tasks. Rather than relying on a single monolithic prompt, it transforms vague requests into structured execution plans with defined phases, roles, and boundaries — acting as an operating system layer for agent logic.

The core problem it addresses is agentic drift: the tendency of LLM-based agents to duplicate work, produce unverifiable outputs, or break down during handoffs between agents. This skill counters that by enforcing ownership rules, explicit exit gates, and evidence-based verification. Every output claim is tied to specific evidence such as logs, diffs, or citations before a workflow phase can proceed.

Four workflow patterns are supported. Discovery-Execute-Verify handles tasks where initial state is unknown. Approval-Gated is designed for high-risk, destructive, or financial operations that require explicit sign-off before proceeding. Parallel Slice manages independent sub-tasks with disjoint ownership to avoid conflicts. Failure-Repair Loop provides systematic debugging support when a process fails and needs structured recovery.

The skill produces a comprehensive Workflow Plan artifact containing a mission statement, complexity classification, a role and purpose matrix, ownership maps covering files and decisions, and concrete verification plans. It is not a runtime execution engine — it produces plans and structures, which an agent or orchestrator then follows. This distinction matters when evaluating whether it fits a given architecture.

Use cases

  • Structure a multi-agent coding pipeline with explicit file ownership and handoff gates
  • Design an approval-gated workflow for agents performing destructive database operations
  • Create a parallel execution plan for independent sub-tasks with defined ownership boundaries
  • Build a failure-repair loop for an agent pipeline that needs systematic debugging on errors
  • Generate a verification plan that ties every agent output claim to logs or diffs before proceeding
  • Classify task complexity and assign role/purpose matrices before dispatching sub-agents

When to use it

  • Orchestrating multiple agents that each own distinct parts of a task
  • Running high-risk operations (financial, destructive, or irreversible) that require explicit approval gates
  • Debugging recurring failures in multi-step agent pipelines
  • Tasks where verifiable evidence of each step's output is required before proceeding
  • Projects where agentic drift — duplicated work or failed handoffs — has been a recurring problem

When not to use it

  • Simple single-agent, single-step tasks that do not involve handoffs or parallel work
  • When a runtime execution engine is needed — this skill produces plans, not executed actions
  • Workflows requiring real-time tool invocations, as this skill has no listed tools
  • When a lightweight prompt is sufficient and the overhead of a full workflow plan is not justified