Skip to content
Agents draft, clinicians sign — no agent path writes to the record alone
LegosphereHealth

The agents

The interesting question is not what an agent can do. It is what it is allowed to do.

Every agent in this system is declared with a tier, and the tier decides its permissions rather than its ambitions. One tier drafts and stops. One acts, reversibly, into a queue a human owns. And the third is not an agent at all — it is the set of computations a language model is never allowed to perform.

Talk to usThe sign-off machine

Three tiers

Permission, not capability

What an agent is allowed to do is declared with the agent rather than decided per feature. The registry will not accept one without a tier, which is what stops the question being answered late and quietly.

Tier A

Touches the record

Drafts only. A clinician signs, or it does not exist.

Anything that would enter the clinical record, or that involves clinical judgement, is Tier A. It produces a draft into a store that is not the chart, a critic gates that draft before a human is shown it, and a licensed clinician signs. There is no configuration that turns this into an automatic write, because a configuration that does is the one that will be on in the busiest week of the year.

  • The draft lives outside the record until it is signed
  • A critic gates the draft before anyone reads it
  • Sign-off is recorded against a named clinician and a timestamp
  • A rejected draft is absent from the assembled record, not merely hidden
Tier B

Operational and administrative

Acts — but every action is reversible and lands in a queue.

Scheduling, waitlists, reorder drafting, referral routing, claim assembly from codes a human already confirmed. These agents genuinely act, because requiring a signature to fill a cancelled appointment slot would mean nobody uses the system. What makes that safe is that the action is reversible and visible: one review and override queue, across every agent, that a human owns.

  • One override queue for the whole roster, not one screen per agent
  • Every action reversible, with the state it changed recorded
  • No Tier-B path may write a clinical resource — that is Tier A by definition
  • An override is data: it is stored, attributable, and reportable
Tier C

Deterministic — and therefore not an agent

The model may call it and explain it. It never performs it.

Drug–drug and allergy interaction checking, dose ranges, controlled-substance refill limits, bed assignment, charge maths and reorder arithmetic are engines with defined inputs and outputs. They are exposed to the roster as tools. A language model that computes a dose is a language model that will one day compute a fluent, confident, wrong one, and no amount of prompt discipline changes that.

  • Interactions, allergies and dose ranges — computed, and hard-blocking
  • Charge and tariff maths — computed, never inferred from a description
  • Bed assignment and eligibility — rules the flow agent must honour
  • Outputs frozen by golden snapshots, so a rule change is visible in review

The safety spine

Preliminary, then signed — and no path skips it

This is the single most important behaviour in the product, so it is one machine that every Tier-A artifact passes through rather than a convention each feature implements for itself.

  1. 01

    The specialist drafts

    The agent produces its artifact — a note, a summary, a set of proposed codes, a discharge packet, a handoff — into the draft store. Nothing about that store is the clinical record.

  2. 02

    The critic gates

    Every Tier-A output passes a critic before it can reach a caller. A draft that fails the critic never becomes something a clinician has to notice was wrong.

  3. 03

    It is preliminary

    A surviving draft is marked preliminary and shown to the clinician. Preliminary cannot be exchanged: it does not leave the building, and it does not appear in an export.

  4. 04

    A clinician signs, and it becomes final

    Sign-off moves the artifact into the record, recorded against a named person and a time. Rejection ends it — and the assembled record provably does not contain it.

The property that matters is negative, so it is tested as a negative: assemble the record after a draft has been rejected or has failed the critic, and the draft is absent from the bundle. Not filtered from a view, not marked inactive — absent. A safety guarantee you can only see by trusting a screen is not a guarantee.

The roster

Every agent, with the tier it is registered under

Grouped by the part of the hospital it works in. The tier beside each name is the useful column — it is what the agent may do, rather than what it is good at.

Documentation and coding

The clinical core, and the founding proof that a drafting agent can be useful without ever writing to a chart.

  • Scribe

    Tier A

    Drafts the encounter note from the consultation. It is the reason the sign-off machine exists, and it was the first thing built against it rather than the first thing retrofitted onto it.

  • Coding

    Tier A

    Proposes what was treated in plain language. Only the terminology adapter turns that into a code — the model never emits a billing code directly, because a fluent wrong code is a claim that pays and an audit that follows.

  • Records summary

    Tier A

    Reads across the record, dedupes and cites. Read-only by construction: it summarises what is there and points at it, and a summary sentence with nothing behind it does not survive the critic.

Front office and patient flow

Operational work where requiring a signature would mean the queue never moves.

  • Registration and matching

    Tier B

    Drives intake and duplicate detection against the matching adapter. Merges remain reversible and ledgered, so an agent-proposed merge is as walkable-back as a human one.

  • Scheduling and waitlist

    Tier B

    Predicts non-attendance and fills released slots from the waitlist. Every fill is reversible and lands in the override queue.

  • Patient flow and beds

    Tier B

    Watches admission, transfer and discharge readiness, and flags bottlenecks. It must honour the deterministic bed-assignment rules; it recommends within them rather than around them.

Clinical support

Closest to the patient, therefore entirely Tier A, therefore entirely draft-and-sign.

  • Results routing and critical values

    Tier A

    Routes results and flags critical values to the right clinician. It flags — it does not interpret. The distinction is the whole design: getting a number in front of the right person quickly is useful, and telling them what it means is not this software’s job.

  • Nursing handoff

    Tier A

    Drafts the structured handoff at shift change. Signed by the nurse handing over, because a handoff nobody signed is a handoff nobody owns.

  • Discharge coordination

    Tier A

    Assembles the discharge packet and the outbound referrals. Follow-up appointments are booked on sign-off — the booking is downstream of the signature, not concurrent with the draft.

Revenue cycle

Assembly and paperwork, on top of decisions a human already made.

  • Claims assembly

    Tier B

    Builds the claim from codes that were already confirmed by a person. It is an assembler, not a coder — the sequence matters, because reversing it is how autonomous upcoding happens.

  • Prior authorisation

    Tier B

    Drives the authorisation exchange and tracks its status, with eligibility checking as a deterministic tool rather than something inferred from a payer’s wording.

  • Denials and appeals

    Tier B

    Reads the remittance, categorises the denial and drafts the appeal for a human to send. The draft is the deliverable; sending is a person’s act.

  • Charge integrity

    Tier B

    Flags missing and mismatched charges before the bill goes out, against the deterministic charge rules rather than against a sense of what looks unusual.

Pharmacy

Where the deterministic engines do the safety work and the agent does the paperwork.

  • Inventory and procurement

    Tier B

    Drafts purchase orders when stock crosses a threshold. The threshold arithmetic is a tool, not a judgement — the agent writes the order, it does not decide the number.

  • Medication reconciliation

    Tier A

    Reconciles the medication list across transitions of care and presents the differences. A pharmacist signs, because an unsigned reconciliation is a list of changes nobody agreed to.

Access and governance

The public-facing edge, and the agents that audit the rest of the roster.

  • Triage and intake

    Tier A

    Collects intake and, on any red-flag pattern, escalates to a human and says so. It is strictly non-diagnostic: it does not tell a patient what is wrong, and it does not tell them it is probably nothing.

  • Patient communication

    Tier B

    Consent-aware outbound messaging across the channels a patient agreed to. Non-diagnostic, and gated on the consent record rather than on a marketing preference.

  • Inbound referrals

    Tier B

    Triages and routes referrals arriving from outside, so a referral does not sit in an inbox while a patient waits for someone to notice it.

  • Compliance and reporting

    Tier B

    Finds documentation gaps, assembles quality and regulatory reporting, and drafts rosters. Including, deliberately, red-teaming the sign-off and override boundaries the rest of this page describes.

The platform underneath

One graph, one registry, one path to a model

  • A director over a registry, not a switch statement

    Each specialist is registered with its name, its tier, its description, its safety rule and its handler. The director routes over that registry, which means adding an agent is a registry entry rather than a rewiring — and it means every agent in the system has a declared tier, because the registry will not take one without.

  • A single path to a model provider

    Exactly one module talks to a model. Every agent goes through it, so the provider, the routing and the residency decision are made in one place rather than in nine. It also means the default in development and test is a deterministic offline stub — the test suite does not depend on a model’s mood.

  • A cost and audit ledger per invocation

    Every agent call writes tokens, latency and cost, and every agent action writes an audit event. Cost per encounter is therefore a number that can be hand-checked rather than estimated, and “what did the agents do to this patient’s record” has an answer.

  • A queue, not an in-process call

    Agent work crosses a durable queue with confirmed delivery. A worker restart does not lose a draft, and a burst of encounters does not become a burst of latency inside a clinician’s screen.

Deliberately not built

Five things absent on purpose

A list of what a system does not contain is more informative than a list of what it does, because the absences are the ones nobody advertises.

  • No diagnostic engine. Nothing here proposes what is wrong with a patient.
  • No treatment-recommendation path. Nothing here proposes what to do about it.
  • No autonomous coder or biller. Codes are proposed for a human to accept and claims are assembled from what that human confirmed.
  • No agent path that writes to the clinical record without a clinician’s signature. Not behind a flag, not for a pilot, not for a demo.
  • No second identity system and no bespoke payment gateway. We verify tokens rather than minting them, and we confirm payment references rather than processing them.

Bring your clinical safety officer

This page is written for the person whose job is to find the path where an agent writes to a record without a signature. If there is one we have not thought of, that is the most valuable thing anybody could send us.

Talk to usThe standard

Get in touch

Talk to the people building it

No chatbot and no ticket queue. Tell us what your hospital or practice actually looks like — how many beds, which departments, what the doctors are doing twice, what your billing team keeps chasing — and someone who works on the software will reply.

info@legosphere.com

Please keep patient identifiers, clinical details and credentials out of this box — it is an ordinary enquiry form, not a channel for protected health information.