# Agent Plan Writing > Agent plan writing is the discipline of designing execution plans for AI agent workflows. The harness matters more than the model. Agent completion rates depend more on action-space design, context en Parent: [AI Agent Ecosystems](https://llms-explorer.com/tree/ai-agent-ecosystems/) · 8 facets · 37 facts · page: https://llms-explorer.com/tree/agent-plan-writing/ ## Overview - Agent plan writing is the discipline of designing execution plans for AI agent workflows. The harness matters more than the model. Agent completion rates depend more on action-space design, context engineering, and orchestration patterns. — source: `https://llms-explorer.com/sources/mdb-context-hub/agent-plan-writing/#overview` ## Output format - When this skill activates, produce a markdown agent plan containing: — source: `https://llms-explorer.com/sources/mdb-context-hub/agent-plan-writing/#output-format` - Workflow Overview - what the system does, which orchestration pattern, and why — source: `https://llms-explorer.com/sources/mdb-context-hub/agent-plan-writing/#output-format` - Agent Roster - each agent's role, model, tools, and context scope — source: `https://llms-explorer.com/sources/mdb-context-hub/agent-plan-writing/#output-format` - Orchestration Graph - how agents coordinate — source: `https://llms-explorer.com/sources/mdb-context-hub/agent-plan-writing/#output-format` - Context Budget - token allocation per agent — source: `https://llms-explorer.com/sources/mdb-context-hub/agent-plan-writing/#output-format` - Safety Constraints - permission boundaries, output validation, human-in-the-loop gates — source: `https://llms-explorer.com/sources/mdb-context-hub/agent-plan-writing/#output-format` - Evaluation Plan - what to trace, quality metrics — source: `https://llms-explorer.com/sources/mdb-context-hub/agent-plan-writing/#output-format` - Failure Handling - per-pattern failure modes and recovery strategies — source: `https://llms-explorer.com/sources/mdb-context-hub/agent-plan-writing/#output-format` ## Orchestration patterns - Five patterns dominate production agent systems: — source: `https://llms-explorer.com/sources/mdb-context-hub/agent-plan-writing/#orchestration-patterns` - Fan-Out: Parallel execution of independent subtasks. Coordinator dispatches to N agents simultaneously. — source: `https://llms-explorer.com/sources/mdb-context-hub/agent-plan-writing/#orchestration-patterns` - Pipeline: Sequential chain where each stage requires the prior stage's output. — source: `https://llms-explorer.com/sources/mdb-context-hub/agent-plan-writing/#orchestration-patterns` - Supervisor: A supervisor agent decomposes the task, delegates to specialists, and synthesizes results. The 2026 production default. — source: `https://llms-explorer.com/sources/mdb-context-hub/agent-plan-writing/#orchestration-patterns` - Debate: Multiple agents reason independently, then argue toward convergence. — source: `https://llms-explorer.com/sources/mdb-context-hub/agent-plan-writing/#orchestration-patterns` - Swarm: Dynamic spawning of agents based on workload. — source: `https://llms-explorer.com/sources/mdb-context-hub/agent-plan-writing/#orchestration-patterns` ## Context window budget planning - Agents consume ~7x more tokens than standard chat sessions. Plan token budgets explicitly. — source: `https://llms-explorer.com/sources/mdb-context-hub/agent-plan-writing/#context-window-budget-planning` - Budget allocation: — source: `https://llms-explorer.com/sources/mdb-context-hub/agent-plan-writing/#context-window-budget-planning` - System prompt: 500–2,000 tokens. Cached input costs 10–25% of normal. — source: `https://llms-explorer.com/sources/mdb-context-hub/agent-plan-writing/#context-window-budget-planning` - Tool schemas: Each MCP tool adds 100–500 tokens to context. — source: `https://llms-explorer.com/sources/mdb-context-hub/agent-plan-writing/#context-window-budget-planning` - Working memory: Reserve 30–50% of context for conversation/reasoning accumulation. — source: `https://llms-explorer.com/sources/mdb-context-hub/agent-plan-writing/#context-window-budget-planning` - Output headroom: Reserve 15–25% for the agent's response generation. — source: `https://llms-explorer.com/sources/mdb-context-hub/agent-plan-writing/#context-window-budget-planning` ## Where this helps - Scoping a new multi-agent system before writing any code, so orchestration pattern, agent roster, and context budget are decided up front instead of discovered mid-build. — [source](https://llms-explorer.com/tree/agent-plan-writing/) *(AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.)* - Choosing between Fan-Out, Pipeline, Supervisor, Debate, and Swarm orchestration when a task's subtask independence and coordination needs aren't yet clear. — [source](https://llms-explorer.com/tree/agent-plan-writing/) *(AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.)* - Estimating token cost for a proposed agent workflow before committing to it, since agents consume roughly 7x the tokens of a standard chat session. — [source](https://llms-explorer.com/tree/agent-plan-writing/) *(AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.)* - Defining safety constraints and human-in-the-loop gates as a first-class planning artifact rather than retrofitting them after an agent system misbehaves. — [source](https://llms-explorer.com/tree/agent-plan-writing/) *(AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.)* ## Project ideas - Build a plan template generator that scaffolds the agent plan sections (Workflow Overview, Agent Roster, Orchestration Graph, Context Budget, Safety Constraints, Evaluation Plan, Failure Handling) for a new multi-agent project. — [source](https://llms-explorer.com/tree/agent-plan-writing/) *(AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.)* - Prototype a Supervisor-pattern system and compare it against a Fan-Out version of the same task to measure completion rate and token cost differences directly. — [source](https://llms-explorer.com/tree/agent-plan-writing/) *(AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.)* - Build a context-budget calculator that estimates system-prompt, tool-schema, working-memory, and output-headroom token costs for a proposed agent roster before deployment. — [source](https://llms-explorer.com/tree/agent-plan-writing/) *(AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.)* - Implement a Debate-pattern workflow for a task with genuine ambiguity (e.g. code review or content moderation) and measure whether convergence actually improves output quality over a single agent. — [source](https://llms-explorer.com/tree/agent-plan-writing/) *(AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.)* ## Antipatterns - Picking an orchestration pattern (e.g. Swarm) because it sounds sophisticated rather than because the task's actual coordination needs call for it — Supervisor is the 2026 production default for good reason. — [source](https://llms-explorer.com/tree/agent-plan-writing/) *(AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.)* - Skipping the Context Budget section of the plan and discovering token overruns only after the system is running in production. — [source](https://llms-explorer.com/tree/agent-plan-writing/) *(AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.)* - Treating Safety Constraints as an afterthought instead of a planning-time deliverable — permission boundaries and human-in-the-loop gates are harder to retrofit than to design in. — [source](https://llms-explorer.com/tree/agent-plan-writing/) *(AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.)* - Writing a plan that specifies models and tools but never lays out the Orchestration Graph, leaving how agents actually coordinate implicit. — [source](https://llms-explorer.com/tree/agent-plan-writing/) *(AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.)* ## Known issues - The ~7x token multiplier for agent workflows versus chat sessions is a planning heuristic, not a guarantee — actual multipliers vary heavily with tool-call volume and working-memory retention. — [source](https://llms-explorer.com/tree/agent-plan-writing/) *(AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.)* - Prompt caching only reduces system-prompt cost (to roughly 10-25% of normal) when the system prompt is stable across calls; frequent prompt changes forfeit the discount. — [source](https://llms-explorer.com/tree/agent-plan-writing/) *(AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.)* - The five dominant orchestration patterns (Fan-Out, Pipeline, Supervisor, Debate, Swarm) are a simplification — most production systems combine patterns (e.g. a Supervisor that fans out to parallel specialists) rather than using one in isolation. — [source](https://llms-explorer.com/tree/agent-plan-writing/) *(AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.)* - Context budget percentages (30-50% working memory, 15-25% output headroom) are starting guidelines, not fixed limits — they need tuning per task, since a long tool-call chain can consume working memory far faster than conversational exchange does. — [source](https://llms-explorer.com/tree/agent-plan-writing/) *(AI-suggested, synthesized from this pack's existing facts — not extracted from a source document.)*