Ownership Boundary

Agent authority is workflow-progression authority. It decides which bounded role may act next and why execution stops, while preserving—not redefining—the evidence and reasoning records consumed by those roles.

flowchart TD
    change{"Which decision changes?"}
    representation["source representation"]
    retrieval["vector execution"]
    meaning["claim support"]
    workflow["role order, convergence, termination"]
    acceptance["flow authority and replay"]

    change --> representation --> ingest["ingest"]
    change --> retrieval --> index["index"]
    change --> meaning --> reason["reason"]
    change --> workflow --> agent["agent"]
    change --> acceptance --> runtime["runtime"]

Decision table

Change	Owner	Reason
parse another document format into a stable record	ingest	changes source admission
select an ANN backend under a budget	index	changes governed retrieval execution
reject a derived claim with no exact support	reason	changes reasoning verification
introduce critique after summarization	agent	changes role sequence and trace
stop after an oscillating verdict window	agent	changes convergence and termination
reject a complete pipeline because tenant entropy policy was exceeded	runtime	changes final flow authority

Reason-to-agent handoff

Reasoning artifacts can enter role inputs, but their claim kinds, statuses, supports, and findings remain reason-owned facts. Agent may schedule critique or verification, retain role output, and record a veto. It must not turn an unsupported claim into a supported one through orchestration metadata.

Agent-to-runtime handoff

Agent publishes a pipeline result and versioned trace containing definition, configuration, role calls, transitions, convergence, termination, telemetry, and final decision evidence. Runtime decides whether that governed output is acceptable in the larger flow. Runtime must not infer missing role history or upgrade an incomplete trace.

Role and controller boundary

Individual role packages perform local work. The pipeline controller owns lifecycle transitions and stop conditions. Roles cannot override the phase, resume after kernel failure, or finalize the workflow implicitly. This separation is defended by architecture invariants, not convention.

Decision custody through a workflow

flowchart LR
    definition["pipeline definition"]
    controller["controller authorizes transition"]
    role["role returns output or error"]
    merge["shard/merge lineage"]
    gates["judge + validate + veto"]
    stop["convergence or termination"]
    trace["PipelineResult + RunTrace"]
    runtime["runtime acceptance"]

    definition --> controller --> role --> merge --> gates --> stop --> trace
    trace -. "governed handoff" .-> runtime

The controller may act only through the pipeline definition and lifecycle. A role may produce content, findings or failure but cannot authorize its next state. Runtime receives the complete agent outcome and may accept or refuse it; it does not fill missing calls or transitions.

Minimum agent handoff

Field	Custody purpose
input/context and configuration fingerprints	identifies the workload and behavior-bearing configuration
pipeline definition and role implementation versions	identifies eligible roles, transitions and terminal states
ordered transitions and call identities	demonstrates that every role invocation was authorized
outputs, typed errors, retries and fallback decisions	retains success and failure rather than only final content
shard/merge lineage, revisions and warnings	proves that no input disappeared behind aggregation
judgment, validation, critique and veto records	preserves decision gates and targeted artifacts
convergence observations and termination reason	explains why orchestration stopped
epistemic disposition, telemetry and trace-completeness findings	separates execution success from evidence quality
final result and versioned trace identity	permits independent reconstruction and cross-artifact comparison

Runtime needs this complete packet to arbitrate a flow. Final prose, a success flag, or an isolated provider response is not an agent handoff.

Boundary breach examples

Breach	Correct owner and response
a role marks an unsupported claim as validated	reason owns claim status; agent records/refuses the invalid role output
a provider asks to skip critique or jump to completion	agent controller refuses the unauthorized transition
merge drops a failed shard	agent retains the shard and produces partial/failed disposition
runtime summary says success while agent trace records veto	runtime comparison refuses the inconsistent handoff; agent trace is not rewritten
valid workflow output is disallowed for a tenant	runtime owns the rejection and retains the intact agent packet

Ownership test

Ask who must make the decision for the record to be valid. Claim verification points to reason. Selecting and ordering roles, sharding, merging, convergence, veto recording, and trace completion point to agent. Accepting the resulting flow under tenant policy points to runtime.