What is progressive disclosure in AI agents?
Progressive disclosure is a context-engineering pattern in which an AI agent first sees only short descriptions of the skills, tools or documents available, and loads the full content only when a task needs it.
The term comes from interface design, where advanced options stay hidden until a user needs them. For agents the reason is context. Every instruction, tool definition and document loaded up front spends part of the model's attention budget, whether or not the task uses it.
Progressive disclosure keeps the up-front cost to a short index. The agent reads the index, decides what is relevant, and pulls in the detail at the step where it matters.
Where it shows up
Agent Skills
Anthropic's Agent Skills (2025-10-16) preload only each skill's name and description. The full SKILL.md file and any bundled scripts load when the agent decides the skill applies.
Tool search
Anthropic's advanced tool use (2025-11-24) lets an agent discover tool definitions on demand instead of loading every one up front.
Tools as code
In Anthropic's code execution with MCP approach, the agent loads tool interfaces as needed and filters data in a sandbox; one example fell from about 150,000 tokens to about 2,000.
Just-in-time references
File paths, links and saved queries stay in context, and the content behind them loads only when read.
Why it matters for architecture
Anthropic's October 2026 guidance on folding subagents into the main loop found that, for many enterprise agents, one loop with progressively disclosed skills and tools beat orchestrator and subagent designs on quality, cost and latency, with one customer seeing roughly twice the speed. Progressive disclosure is what lets a single agent hold many capabilities without a bloated prompt.
What it depends on
The agent only loads what its descriptions point it to, so the descriptions do the routing. OpenAI's guidance for long-running agents advises writing skill descriptions as routing logic, including negative examples of when a skill should not be used. A vague description means the right detail never loads.
How Motiontic's engineers apply it
Motiontic's engineers keep the always-loaded layer short, move procedures and reference material behind descriptions, and test triggering with evals that include prompts where a skill or tool should stay unused.
Loading everything up front compared with progressive disclosure
| Everything up front | Progressive disclosure | |
|---|---|---|
| Context at the start | All instructions and tool definitions | Names and short descriptions |
| Cost of adding a capability | Grows every prompt | One more line in the index |
| Main risk | Context rot and distraction | The right detail is never loaded |
| What needs care | Pruning the prompt | Writing descriptions that route well |
Common questions
What is progressive disclosure in Agent Skills?
Only each skill's name and description sit in the agent's context. The full instructions and bundled files load when the agent decides a skill is relevant to the task.
How does progressive disclosure save tokens?
It replaces full instructions and tool definitions with short descriptions until they are needed. Anthropic reported one MCP example falling from about 150,000 tokens to about 2,000 when tools were loaded on demand.
Is progressive disclosure the same as RAG?
They are related. Both load information on demand. RAG retrieves passages by similarity search, while progressive disclosure lets the agent choose from a short index of skills, tools or documents it can open.
Related
Skills vs MCP
Connectivity to systems, compared with packaged expertise the agent loads on demand.
Tool search
Load tools when they are needed, not all at once before the work starts.
Context engineering
Curating everything in the model's view, not just the wording of the prompt.
Subagent
A helper agent with a clean context that hands back a short answer.
AI agent development
Agents that do real work for users, with the evals, traces and guardrails to keep them in production.
Sources
- Anthropic, Equipping agents for the real world with Agent Skills (2025-10-16)
- Anthropic, Introducing advanced tool use (2025-11-24)
- Anthropic, Code execution with MCP (2025-11-04)
- Anthropic, Effective context engineering for AI agents (2025-09-29)
- Anthropic, When to fold subagents into the main agent loop (2026-10-22)
- OpenAI, Shell + Skills + Compaction: tips for long-running agents
Running agents that have to hold up?
Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.