tool-design-composition

Tool Composition and Discovery Design

A composable tool set has four properties: consistent input and output shapes, so one tool's output feeds the next tool's input; batch support, so the agent does not loop one item at a time; more than one abstraction level, so the agent picks the granularity the task needs; and discoverability, so the agent finds the right tool without trial and error.

Procedure

1. Map the current set: every tool, its type, and the sequences agents actually run (from traces, not from intent).
2. Find bundle candidates (Task Bundle): sequences repeated across runs with no branching between steps.
3. Find batch candidates (Batch Operation): the same tool called in a loop over a list.
4. Find ladder gaps (Abstraction Ladder): capabilities with only a raw low-level tool and no intent-level tool, or the reverse.
5. Add a mode parameter (Operation Mode) that defaults to preview on every destructive or costly command.
6. Design discovery (Dependency Hint, Tool Registry, Schema Explorer): hints in descriptions, a registry or search tool once the set is large, and layered schema exploration for data sources.
7. Prune: merge near-duplicates and retire tools with no traffic. Fewer, better-described tools beat many similar ones.
8. Verify: run each new tool on its own, then run the agent and read its trace to confirm it picks the new bundle, batch form, or preview mode instead of the old sequence.

Rules with examples

Offer the capability at more than one altitude (Abstraction Ladder)

low   execute_query(sql)                       full control, expert use
mid   search_records(table, filters)           structured, safe
high  find_customers_like(description)         intent-level, does the work

Each rung has a stated use case; higher rungs call lower ones internally; descriptions say when to use which. Do not ship only the low rung and expect the agent to compose it correctly every time.

Bundle the sequence the agent always runs (Task Bundle)

search_users then open_dm then send_message becomes dm_user(name, message). Name the bundle after the task, not the steps. Handle intermediate failures inside the bundle and return a composite result that reports each step. Keep the atomic tools available for edge cases, and say in the bundle's description that it replaces the sequence.

Accept lists (Batch Operation)

Any tool the agent calls per item over a list gets a batch form: an array input, a documented maximum batch size, per-item results with status, and a partial-success report (see tool-design-output). Batch the upstream calls internally where the API allows. Decide up front whether a batch is all-or-nothing or partial, and say which.

Explore before executing (Operation Mode)

Give commands a mode parameter that defaults to the safest option:

explore   list what the tool could do here, read-only
preview   show exactly what would change, no side effects
dry_run   validate inputs without executing
execute   perform the action

Return the same result structure in every mode so the agent can compare a preview to the real run. A schema mode that returns the upstream's complex input shape is useful when the shape cannot be expressed in the tool interface.

Make the sequence explicit when it matters (Tool Chain)

For multi-step processes with a fixed order, define the chain as ordered steps with data passing between them and checkpoints for resume. A chain is a specification the agent follows, or a single orchestrated tool that runs it; either way the order is not left to inference.

Fan out, then merge (Scatter-Gather Tool)

A tool that consults several sources queries them in parallel, merges the results into one canonical shape, and reports per-source status so a failed source degrades the result rather than failing it.

Discovery: how the agent finds the right tool

Say what to call first (Dependency Hint)

Descriptions and parameter docs name prerequisites ("userId: if you only have an email, call search_users first"), follow-ups ("after this, get_user verifies the change"), and alternatives ("if this fails, try send_email"). Repeat the hint in the error the tool returns when the prerequisite is missing.

Catalog the set when it is large (Tool Registry, Capability Matching)

Past roughly twenty tools, tool-selection accuracy drops and every definition costs context tokens. Provide a list_available_tools(category) registry with name, description, category, and auth requirements, and a find_tool_by_intent(intent) search that ranks tools by semantic match with confidence and usage examples. Where your runtime can defer tools, keep the must-use tools always loaded and let the long tail be discovered.

Reveal structure in layers (Schema Explorer)

For data sources, do not dump the full schema. Layer it: list_tables(), then describe_table(name), then sample_rows(table, limit), then get_query_hints(table). Each layer's description names the next.

Check before relying (Health Check)

A fast (sub-second) check_service_health(service) returns healthy, degraded, or unavailable with specific reasons, so the agent does not spend turns on a dead backend. Cache the status briefly.

Across the system

• Tool Gateway: one entry point routes to many backends and aggregates discovery, so the agent sees one consistent set.
• Tool Adapter: wrap a legacy or awkward API in a clean tool: LLM-friendly description, constrained inputs, shaped output. The adapter hides the legacy shape entirely.
• Canonical Tool Model: shared User, Task, Event shapes and field names across every tool, mapped to and from each backend.
• Tool Versioning: version in the name (send_email_v2) or metadata, run versions in parallel during migration, and deprecate with a migration note in the old tool's description.

Anti-patterns

• Thirty atomic tools and no bundle for the sequence 80% of runs perform.
• delete_files with no preview mode.
• Two tools that differ only in one optional parameter.
• A registry that lists tools without their auth requirements or categories.
• A composite tool built before any trace showed the sequence being used.
• A migration that changes a tool's behavior in place instead of versioning it.

On Runtype

• Tool count and residency. Runtype does not cap the number of tools in a dispatch; it caps the request's serialized runtimeTools at 1 MB. A model provider can still refuse a long tool list (TOOL_COUNT_REFUSED_BY_PROVIDER), so use tool search for large tool sets. Tool search is on by default for multi-turn agents: at 20 or more tools, Runtype keeps a hot set loaded and defers the inline runtime tools behind a tool_search tool that the model calls to load more. Saved, built-in, and MCP tools stay loaded. Set tools.toolSearch.enabled: false to send every tool on every request. Single-turn agents and flow prompt steps always send every tool, so keep those sets small or split them across subagents. Either way, make descriptions distinct, because they are what the model chooses by and what tool_search matches against.
• Tool-call budget. An agent makes at most 10 tool calls per execution unless you set config.tools.maxToolCalls. A per-item loop over a list uses up that budget quickly, so ship a batch form, or raise the cap on purpose.
• Task bundle and tool chain. A flow is the platform's explicit chain: fixed step order, data passing, per-step error handling. Expose it as one tool (toolType: "flow") when the agent should run the whole sequence as a single call. Map tool parameters to flow inputs with parameterMapping. The model receives the flow's terminal step output; to return a composite status object instead, assign it to a flow variable and name that variable in outputVariable. A variable name that starts with _ is rejected, and a variable the flow never assigns fails the call.
• Code Mode for batch, chain, and fan-out. When the logic between calls varies by input (loops over a list, conditional chains, several sources merged), set config.tools.codeModeConfig on the agent or prompt step. toolPool lists the tool ids the model can call (wildcards such as mcp:* work), and timeoutMs defaults to 60000. The model writes JavaScript that calls those tools in a sandbox, and intermediate values never enter model context. Choose a flow tool when the order is fixed; choose Code Mode when the model must write the logic.
• Binary results. A binary tool result, such as a screenshot or a generated image, carries a runtype-asset:// handle. To accept it in your own tool, declare the input as a string with "contentEncoding": "base64"; a handle passed there is replaced with the stored bytes before your tool runs, so the model never copies base64. A parameter without that declaration receives the handle string unchanged.
• Abstraction ladder. Built-in and Orthogonal catalog tools are the low rung; custom, external, and mcp tools the middle; flows and subagents the top. Compose subagents as saved subagent tools, inline subagent tools, or dynamic spawning through config.tools.subagentConfig, which gives the agent a spawn_subagent tool. Subagents have their own caps; read get_platform_documentation(topic="subagent-delegation").
• Gateway and adapter. An MCP server attached through config.tools.mcpServers is a Tool Gateway: one connection, many tools. An external tool is a Tool Adapter: it wraps an upstream API behind a clean interface. For a GraphQL API, use an external tool that POSTs the query; the graphql tool type is not executable.
• Operation mode. On an agent, gate destructive tools with config.tools.approval.require (for example, ["delete_files"]; patterns such as mcp:* work). Approval asks a person before the call runs; it complements a preview mode but does not replace it. A tool-call flow step runs a catalog tool deterministically with no approval gate, so there a preview must be a distinct tool or a mode parameter. Approval details: get_platform_documentation(topic="agent-design").
• Description length. A new saved tool's description is capped at 500 characters. When dependency hints, replacement notes, and deprecation notes do not fit, move them into parameter descriptions or the error text.
• Registry and discovery. list_tools covers saved tools, discover_mcp_server_tools an MCP server, and get_platform_documentation(topic="builtin-tools") the catalog. A tool-call step that names a catalog id nobody answers to is rejected as TOOL_CALL_STEP_UNKNOWN_CATALOG_TOOL.
• Traces and verification. Find real sequences with list_runs (filtered by agent_id) and trace_execution. Test one tool with execute_tool, then run the agent with execute_agent and read the trace to confirm the tool choice.
• Versioning. Agents and flows are versioned (publish_agent_version, publish_flow_version); a tool behavior change ships behind a new published version rather than in place.