Metaphor Labs is building an execution control layer that analyzes agent runs to safely reuse repeated work, improve multi-step execution, invalidate stale results when dependencies change, and preserve source-to-output provenance.
Lower model spend and latency. Keep reused work fresh. Explain every result.
A single request can trigger many model, retrieval, and tool calls. As workflows become longer and more dynamic, repeated work, stale dependencies, and incomplete traces become operational problems—not just implementation details.
request A → review this account
request B → assess this customer
equivalent intent → repeated execution
Exact-match caching misses semantically equivalent requests, so models, retrieval, and deterministic steps are often recomputed.
TTL-based caches do not know which source, document, database row, or tool result made an answer valid.
iterate↙waste can repeat inside each cycle
Loops and branches are necessary, but work inside them may still be parallelizable, batchable, or safely reusable.
When a result needs review, teams need the source → retrieval → tool → prompt → model → output chain behind that specific result.
Three connected capabilities help teams reduce repeated work, improve execution efficiency, and understand why each result happened—without replacing their existing orchestration, models, data, or tools.
Recognize equivalent intent beyond exact text. Track the dependencies behind reused results so relevant changes can invalidate them before stale output is served.
Profile workflows as they run, including loops and branches. Extract only safe acyclic regions for parallelization, batching, and deterministic reuse while preserving control-flow behavior.
Preserve source-to-output lineage across retrievals, tool calls, prompts, and model steps so debugging, review, and impact analysis start with evidence.
Metaphor sits alongside agent orchestration and above downstream models, retrieval, and tools. It focuses on application-level execution: what can be reused, what can run more efficiently, when prior results become stale, and how outputs are traced.
As production agents become longer-lived and more multi-step, efficiency and reliability depend on understanding the application context around each call—not only serving inference faster.
One user action can trigger many model, retrieval, and tool calls. Repeated work multiplies quickly.
Reusing prior work is valuable only when the system knows whether the inputs that made it valid have changed.
Teams need to reconstruct why a result happened across data, tools, prompts, and models—not just search logs after an incident.