Lab Agent OS · for regulated labs
A crew of compliance-grade AI agents for deviations, CAPA, SOPs, training and audit prep — every step traced to its source, every decision signed by a person.
Most quality work is assembling evidence that already exists in five systems, in the right format, by a deadline. Lab Agent OS drafts the deviation, scopes the impact, proposes the CAPA and routes it for signature — inside the LIMS and QMS you already run.
21 CFR Part 11
EU GMP Annex 11
GAMP 5 · CSA
SOC 2 Type II
The work between the science
Most quality work is assembling evidence that already exists in five systems. Agents are good at exactly that — and it’s the work nobody on your team will miss.

Today: the analyst writes it up from memory, then QA rewrites it.
Drafts the deviation from the instrument signal, LIMS runs and the governing SOP, with every claim cited
Builds a 5-Whys / Ishikawa from telemetry, maintenance and training records
Proposes CAPA with measurable effectiveness checks and tracks every due date
Autonomy L1: drafts for a person to edit and submit

Today: doc control redlines by hand after every change.
Finds every SOP, form and training module a change touches
Produces tracked redlines and an impact summary reviewers actually read
Flags conflicting SOPs and documents overdue for periodic review
Autonomy L1: drafts for a person to edit and submit

Today: managers cross-check the training matrix; planners juggle fixed intervals.
Assigns read-and-understand on the SOP effective date and chases overdue items
Blocks unqualified analysts from running a method in LIMS
Schedules calibration from predicted drift and books the service window
Autonomy L2: reversible, logged actions

Today: two weeks of binders before every inspection.
Runs a mock inspection every night against your SOPs and regulations
Keeps a live evidence room that answers “show me” requests with source records
Drafts responses to auditor requests during the inspection
Record approval stays human: agents never apply an e-signature
How an agent works
Every agent runs the same governed loop: grounded in your SOPs, acting only through connectors you’ve scoped, checked by a second model, and stopping where a regulated decision is made.

Triggered by an instrument alarm, an OOS result, an SOP change or an overdue CAPA. The agent pulls the effective revision of every SOP, spec and prior record that governs the event, then plans against them.

Queries and updates LIMS, ELN, QMS and LMS through scoped, least-privilege connectors. No copies, no parallel system of record, nothing to migrate.

An independent checker model resolves every citation, tests the draft against acceptance criteria and sends weak work back. Low confidence goes to a person, not forward.

The record arrives complete, cited and risk-ranked. Approval is a Part 11 e-signature by an authenticated human. Reviewer edits feed the agent’s evaluation set — never silently into the live model.
The crew
Agents hand work to each other — a deviation spawns a CAPA, a CAPA triggers an SOP change, the SOP change triggers training — so one event is followed all the way to closure.

Turns alarms, OOS/OOT results and analyst notes into a cited deviation draft.
Reads: Trucision instrument alerts, LIMS results and run history, analyst notes
Produces: deviation record, affected lots and samples, ICH Q9 preliminary risk
L1 draft · 21 CFR 211.192 · ISO 17025 §7.10

Runs the investigation workbench and chases CAPA to closure.
Reads: open deviations, instrument health history, training and maintenance logs
Produces: root-cause analysis with evidence, CAPA plan and owners, effectiveness checks
L1 draft · ICH Q10 · ISO 13485 §8.5.2 · ISO 17025 §8.7

Maps every document, validation and training record a change touches.
Reads: change requests, equipment and method master data, validation inventory
Produces: impact assessment, revalidation checklist, linked SOP redline tasks
L1 draft · EU GMP Annex 15 · ICH Q10

Keeps controlled documents coherent and current.
Reads: controlled document library, approved change requests, regulatory updates
Produces: tracked redlines, conflict and gap report, periodic review queue
L1 draft · 21 CFR 211.100 · ISO 13485 §4.2.4

Assigns, tests and enforces analyst qualification.
Reads: role and training matrix, SOP effective dates, LIMS method assignments
Produces: assignments and reminders, competency checks generated from the SOP, qualification flags
L2 reversible · 21 CFR 211.25 · ISO 17025 §6.2

Condition-based schedules, built on Trucision Predictive Maintenance.
Reads: predicted drift and health scores, sample queue, service contracts
Produces: condition-based schedule, vendor bookings, equipment log entries
L2 reversible · 21 CFR 211.68 · ISO 17025 §6.4

A mock inspection every night, and a live evidence room.
Reads: all quality records, regulations and sponsor SOPs, past findings
Produces: daily readiness score by area, evidence packets, draft responses
L1 draft · ISO 17025 §8.8 · FDA inspection readiness

Checks CoAs against spec and tracks supplier qualification.
Reads: CoAs and incoming inspection, supplier master, material specs
Produces: CoA check results, requalification calendar, draft SCARs
L1 draft · 21 CFR 211.84 · ISO 13485 §7.4
Governed by design
Bounded autonomy you set per workflow, a validation package in line with GAMP 5 and Computer Software Assurance, and a record an inspector can follow step by step. In critical GMP decisions, AI assists and people decide.

Start every agent in shadow mode. Promote it one level at a time as its record earns it.
L0 Shadow: runs silently beside your team; output is compared, never shown
L1 Draft · L2 Act reversibly · L3 Close routine items within pre-approved rules
Never: sign, approve or release. Hard-coded, not a setting

An agent has a configuration, an operating range and a performance record. We deliver the evidence with it.
IQ: model, prompts, tools and scopes hashed and locked under change control
OQ: replays hundreds of your closed deviations and CAPAs against agreed criteria
PQ: 30 days in shadow on live work. Drift past a threshold drops the agent a level

“How did you conclude that?” — open the run. Every step is kept as an ALCOA+ record.
Agent ID, model version and approving user on every record
Every claim links to the exact SOP revision, LIMS result or telemetry window
Append-only audit trail; reviewer edits shown as diffs against the draft

No new system of record. Records stay in your validated QMS and LIMS.
Connects to QMS, LIMS/ELN, document control, LMS, ERP and Trucision telemetry
Private deployment; your records are never used to train foundation models
Pause any agent, site or workflow instantly; work in flight returns to a human queue
30-day pilot
Pick one workflow. In 30 days you’ll have a shadow-mode agent working on your real records, and the validation evidence to decide whether it goes live.

Choose the workflow, map SOPs and connectors, and agree acceptance criteria with QA.

Run the agent on your closed records and review the scored results together.

Run at L0 on live work, trending agreement and reviewer edits every day.

You get a validation summary and a go/no-go for promotion to L1.