TracePact

Core thesis

Agentic latency is not primarily a “slow model” problem. It is a coordination problem among typed effects.

Treat every model call, tool call, retrieval, guardrail, handoff, cache lookup, UI update, human approval, retry, and state rehydration as an effectful operation with a contract:

• What information or authority does it require?
• What information or authority does it produce?
• Does it have to run fresh, or is replay/cache acceptable?
• Is it copyable, replayable, affine, linear, or ephemeral?
• Is it read-only, externally mutating, user-visible, or safety-critical?
• Which later spans actually consume its result?
• Could it be routed to a different provider, collapsed into another span, run in parallel, run speculatively, moved off the critical path, or replaced by deterministic code?

The skill’s job is to compile the observed run into a Latency Treaty IR: a machine-readable execution contract plus a counterfactual rewrite plan. The output should not merely say “parallelize tools.” It should state which dependency, policy, authority, replay, or quality invariant makes the rewrite legal.

What makes this skill different