Orcher · Orchestration
Many agents. One accountable human.
Intake, verification, finance, and audit agents can each be correct while the outcome is indefensible. Orcher holds the directive constant across every hand-off so the chain never breaks.
Vol. I · p.16, p.17


Worked example
Claims-to-cash, end to end
Four agents, four systems, one adjuster whose authority is exercised at every step. If any Layer 1 check fails, the directive halts and nothing partial is written.
01
Intake agent
Parses the submitted claim, normalizes it, and attaches the directive issued by the named adjuster.
02
Verification agent
Reconciles coverage, eligibility, and prior history against the policy system of record.
03
Finance agent
Prepares the payment instruction — and cannot commit it until the Logic Scrubber clears the write.
04
Audit agent
Assembles the evidence set and closes the cycle with one hashed record in the Immutable Audit Ledger.
Clinical prior authorization runs the same shape: an intake agent assembles the request, a clinical-criteria agent reconciles it against policy and chart data, a utilization agent prepares the determination, and the reviewing clinician's role — not a generic service account — is the authority under which the determination is issued.
Why coordination fails
Fragmented authority is the failure mode.
Naive multi-agent configurations fail between 41% and 87% of the time. Coordination research using shared decision state raises agreement from roughly one third to 93% — but agreement is not authorization.
41–87%
failure rate of naive multi-agent configurations
~⅓ → 93%
shared-decision rate with coordinated state
150+
organizations behind A2A v1.0 delegation semantics
1
hashed record closing the whole chain
One directive
Immutable across every agent and every hand-off.
One role
The human authority being exercised, cryptographically bound.
n agents
Each scoped to that role, never to its own standing credential.
One record
A single Verified Execution Cycle, not four disconnected logs.

Orchestration without attribution is just latency.
Orcher makes the hand-off the safest point in the workflow rather than the weakest.
