CONCEPT Cited by 1 source
Progressive tool disclosure¶
Definition¶
Progressive tool disclosure is the practice of exposing an LLM agent's available tools on demand rather than loading every tool schema into context upfront. The agent sees a compact summary of what's available; full schemas are fetched only when the model reaches for a specific tool.
Atlassian Long Horizon implements this with two meta-tools per
product namespace:
- {product}__get_tool_schema โ returns the full input schema for
one specific tool on demand; its description carries a compact
one-line summary of every tool in that namespace.
- {product}__invoke_tool โ executes a tool in that namespace by
name with arguments.
Frequently-used tools (search, todo-list, file I/O, memory retrieval) remain flat at the top level because they're called nearly every turn and are worth resident schema cost. (Source: sources/2026-06-18-atlassian-long-horizon-reasoning-engine)
Why it helps¶
- No per-product schema tax. The old architecture re-paid each product's full schema bundle every time its sub-agent was invoked. Progressive disclosure means only specifically-reached tools pay schema cost.
- Scales with tool count. Adding a new product or new tools within a product doesn't inflate per-iteration context cost.
- Bounded discovery cost. Schemas are fetched once per tool per
task (not per iteration). SKILL.md often lets the model go straight
to
invoke_toolwithout a separate schema fetch.
Trade-offs¶
- Introduces a discovery step: for unfamiliar tools, the model first
calls
get_tool_schematheninvoke_tool(2 calls vs 1 for a flat tool). - Requires well-authored one-line summaries in the meta-tool description for the model to select correctly.
- SKILL.md maintenance burden per namespace.
Relation to tool-surface minimization¶
patterns/tool-surface-minimization focuses on reducing the number of tools visible. Progressive disclosure is complementary: it keeps many tools available but only pays schema cost for the ones actually used. Both aim to keep the context budget tractable.
Seen in¶
- sources/2026-06-18-atlassian-long-horizon-reasoning-engine โ canonical articulation in Atlassian Long Horizon.