LLM Wiki — Synaptic Lattice

Capability as Method Calls

Properties1
tagsprompt-engineering, architecture-pattern, token-usage-reduction

Leitura de 2 min

Capability as Method Calls

Overview (confidence: medium)

Capability-as-method-calls is an architecture where agent capabilities are exposed to the model as typed method calls in agent code (agent.searchGoogle(), agent.sendEmail()) rather than as tool schemas advertised in the context window. The model invokes capabilities by calling methods directly; the infrastructure layer — rate limiting, retries, auth, execution — runs outside the prompt entirely.

This is a third answer to tool-schema bloat, distinct from the client-side approaches in tool-context-management (which decide when and whether schemas load) and from the gateway pattern in code-mode-orchestration (which moves governance into a sandbox but still reads tool stubs on demand). Capability-as-method-calls removes the schema payload from the model’s context altogether, keeping the model focused on reasoning and decision-making.

Key Points

  • Context cost: MCP tool schemas are re-sent on every message with no cross-turn caching — a typical multi-server setup carries 12,000–20,000 tokens/turn before the first query. Method calls sidestep this by never injecting the schema at all.
  • Where the model’s knowledge comes from: with no schema in context, the model must already know the method names and signatures, or discover them another way — the architecture trades prompt-space for documentation/code knowledge.
  • Execution separation: capabilities that are heavy on retries, auth, or side effects (web search, email, image generation) run in the execution layer, so their failure modes don’t consume model reasoning cycles.
  • Fits capability-heavy, execution-light-reasoning workloads: keep Claude Code for reasoning and coding while offloading broad capability sets to an agent infrastructure that doesn’t inflate token counts.
  • Vendor origin: documented by MindStudio as its Agent Skills Plugin (npm @mindstudio-ai/agent) — vendor-self-published, no independent benchmark, so treat the savings claims as directional.

Implications

This is an architectural alternative, not a configuration tweak. It moves tool governance from the prompt (client-side approaches) or a sandbox (gateway) into the application code itself. For token-usage-reduction, it is the strongest version of “keep schemas out of context” — but it requires the model to hold capability knowledge another way, so it suits agents with a small, stable, well-known capability set rather than open-ended toolsets.

Open Questions

  • Whether method-name knowledge scales: does the model reliably call the right method as the capability count grows, without schema discovery in context?
  • Whether the absence of schema descriptions degrades correctness on unfamiliar or infrequently used capabilities.
  • Whether the savings hold under independent measurement (currently vendor-only claims).

tool-context-management code-mode-orchestration token-usage-reduction claude-code-system-prompt bifrost

Sources

  • raw/prompts/articles/mindstudio-claude-code-mcp-token-overhead.md

Visão de gráfico

Backlinks