Agent Harness Wiki

2 min read

Partial-Order Trace Semantics

Definition

Partial-order trace semantics treats a run not as one total sequence of events, but as a structure in which some events are causally ordered, some are concurrent, and some represent branching choice. For agent harnesses, this matters whenever parallel subwork, review gates, or multi-agent coordination are collapsed into a mere list of turns.

What the new papers add

Wang argues that structural operational semantics can be extended from ordinary labelled transition systems to pomset-style transitions, so the semantic unit becomes a partially ordered multiset of actions rather than one token at a time. Edixhoven et al. add the other half of the problem: concurrency alone is not enough, because choices also need compact representation. Their branching pomsets encode both partial order and branching behaviour for multi-agent choreographies.

Harness takeaway

The harness analogue is straightforward:

  • subagents and teams create concurrent work, not just longer transcripts
  • review and approval introduce causal barriers
  • delegation choices create branches, not only longer serial histories
  • critical-path reasoning is therefore better expressed over dependency structure than over raw event count

This sits directly beside orchestration-topologies, work-management-primitives, and automation-and-background-work. A work graph is already an admission that the transcript is not the right semantic object.

Why this matters for current harness theory

A scheduler lower bound is a good first move, but it is still only a scalar shadow of a richer object. The next serious step would be to treat harness traces as partial orders or pomset-like structures, from which critical path, branching, and realizability properties can be derived. That would make the current scheduler layer look like a coarse quotient rather than an isolated trick.

Open questions

  • Which harness events should remain atomic, and which should be grouped into a concurrent action set?
  • Is pomset structure enough, or do harness traces eventually want full event structures?
  • How should branching for retries, evaluator disagreement, or tool failure be represented?
  • How should this layer interact with formal-methods-for-agent-harnesses and probabilistic-epistemic-updates?

Read this with orchestration-topologies, work-management-primitives, automation-and-background-work, formal-methods-for-agent-harnesses, probabilistic-epistemic-updates, and new-harness-design-notes.

Graph View

Backlinks