Core work system · Basis 29 August 2026

02 / 10
Implemented prototype

Ordivon Harness

Preserve the structural truth of one Agent Run while leaving semantic cognition and action choice with the Agent and Task/domain meaning with the caller.

The independent durable cognitive execution substrate for bounded Agent Runs. Harness preserves exact structural cognition, action, and effect continuity while the Agent chooses semantic cognition/action and the caller/domain retains Task meaning and final acceptance.

Projected status
Operational pre-1.0 engineering prototype with a supported caller-facing Agent Run surface
For
Agent builders who need durable bounded execution, explicit cognition/action authority, exact Provider/Tool recovery, and caller-owned completion semantics without a Host or generic Memory/RAG dependency.
Evidence for status
Current Status and Architecture freeze six distinct state domains—canonical history, durable cognition, interaction cognition, attempt cognition, execution control, and effects—while HarnessAgentRun provides the supported Python composition surface. Recent verification work binds typed evidence to creation lineage and fails closed on structured-result conformance without promoting semantic evaluation or Provider continuation into Harness.

Question this project must answer

Which structural facts must survive interruption so an Agent can resume from exact current cognition and authority without Harness choosing relevance, world truth, or Task completion?

Current operating model

01Canonical bounded-Run history and immutable Run authorityprototype
02Durable WorkingSet/WorkingView cognition mechanicsprototype
03Interaction and attempt-local cognition reconstructionprototype
04Request-bound Harness/Tool action authorityprototype
05Provider/Tool effect fencing, recovery, receipts, and CompletionProposalprototype

Standing in this projection

Operational pre-1.0 engineering prototype with a supported caller-facing Agent Run surface

Operational pre-1.0. One independent Harness Journal/CAS owns bounded Run continuity; Agent-selected cognition is distinct from history, current action authority is request-bound, ambiguous Provider/Tool effects are recovered conservatively, and HarnessAgentRun is the supported normal Python composition surface. Generic Memory/RAG, semantic ranking, Provider-opaque continuation, Strategy scheduling, and domain completion remain outside Harness.

Capability evidence

6separate Run state domains
HarnessAgentRunsupported Python composition surface
owner-externalsemantic relevance and final acceptance

What this project owns

Facts, products, or methods maintained here.

  • Canonical bounded-Run history and immutable Run authority
  • Durable WorkingSet/WorkingView cognition mechanics
  • Interaction and attempt-local cognition reconstruction
  • Request-bound Harness/Tool action authority
  • Provider/Tool effect fencing, recovery, receipts, and CompletionProposal

What it leaves elsewhere

Responsibilities and claims outside its boundary.

  • No generic Task/domain database or final acceptance
  • No Runtime Workspace/Job/Attempt ownership
  • No semantic ranking, Memory/RAG truth, or automatic knowledge promotion
  • No generic Provider-opaque continuation or Mandate scheduler

Questions and current judgments

01

answered

Which Harness objects survive live Provider replacement without duplicating Host or Runtime state?

02

answered

Does the caller-neutral Harness boundary remain valuable across different workloads and execution strategies?

03

answered

Which Provider-required state must survive a Tool turn without becoming Agent cognition?