Progressive disclosure

concept · updated Jul 15, 2026

person concept tool org talk claim — click a node to jump to its page; hover an arrow for the relation

A design principle in AI agent engineering where instructions, context, or capabilities are surfaced to a model incrementally and at the moment of need, rather than being front-loaded all at once at the start of an interaction.

Core Principle: Just-in-Time Context Delivery

Ryan Lopopolo (OpenAI) advocates for progressive disclosure as a foundational property of well-built agent harnesses. His argument is that overwhelming an agent with all instructions upfront is counterproductive — instead, a good harness should defer or surface instructions precisely when they become relevant: "figuring out ways to either defer or just in time surface those instructions is kind of what a good harness should do." This framing treats progressive disclosure not as a nice-to-have but as a defining characteristic of harness quality.

Role in Agent Skills Architecture

Philipp Schmid (Google DeepMind) identifies progressive disclosure as the distinguishing structural property of agent skills specifically, contrasting them with other agent design patterns: "the big difference with skills is that they work on progressive disclosure." This positions progressive disclosure as not merely an implementation detail but an architectural principle that defines what a skill is — capabilities that emerge and become available to the agent contextually, tied to the evals-driven development process Schmid advocates in the same talk.

Synthesis

Both speakers treat progressive disclosure as part of avoiding context pollution and reducing cognitive/computational load on the model at any given step. In harness engineering it manifests as deferred instruction injection; in skills architecture it manifests as capability gating. The shared intuition is that agents perform better when their working context is scoped to what is immediately actionable.