Autonomous-Lab Experimentation Agent (Plan to Run Record)
Plan experiment campaigns, schedule them against instrument and reagent availability, run them inside an envelope your lab head approved, watch them and stop them — the run recorded, no result released as reportable.
Take the campaign objective, the envelope it may run in and what the last round returned.
02
Read instrument state, calibration status, reagent lots and the consumables on the deck.
Reason
03
Propose the next round of conditions from what came back, inside the ranges the envelope allows.
04
Schedule runs against instrument availability, reagent stability and what may not share a deck.
05
Build the run as executable steps and test each precondition before anything is asked to move.
Decide
06
Hold anything outside the envelope, and work no safety or biosafety owner has authorised.
07
Stop a run on an alarm, a failed check or a deviation, and park the instrument safely.
Out
08
Write the run record — what ran, on what, with which materials, under which plan version.
09
Return the output as run data for a named scientist, never as a reportable value.
→Product statement
The agent plans, schedules and runs inside the approval boundaries agreed during implementation. The envelope, authorisation for hazardous, controlled or biological work, and the release of a reportable value stay with the lab head, the named scientist and your biosafety and safety owners.
Example workflow
One campaign round, end to end
AgentHuman
1Round requestedA campaign objective, the envelope it may run in and what the last round returned
2Preconditions readInstrument state, calibration status, reagent lots and expiry, tips, plates and the deck as it stands
3Conditions proposedThe next set of conditions, each inside the ranges the envelope allows, with what the round is meant to settle
4Run built and scheduledExecutable steps, instrument and reagent bookings, and the rules that keep two things off one deck
No human action required
Stages 1 to 4 run without a person in the loop — nobody is asked to look until the preconditions have been read and the round is built. A condition the envelope does not cover stops the round there, before an instrument is booked.
5DecisionSplits on the envelope check and the confidence gate
Inside the envelope
Enters the queue as an approved run.
Outside it, or unauthorised
Held with the offending condition marked.
Lab head, named scientist and safety owners
Read the proposed conditions, the deck, the hazards and whatever falls outside the envelope.
Approve · Amend · Send back
Approved — handed back▼
6Run executed and watchedStarted only where the envelope, the authorisation and the instrument allow it; stopped on an alarm or a failed check
7Outcome evaluatedAborted runs, preconditions that were wrong, conditions the scientist amended and what did not reproduce
Amendments
Conditions a scientist changes before a run are counted in the evaluation.
What should not run autonomously
Human approval stays in control
Outside the boundary — human approval required8 items
Running anything outside the approved envelope.
Authorising work with a hazardous, controlled or biological agent.
Changing an instrument's safety configuration or a calibration.
Overriding an interlock, an enclosure or an alarm.
Automation boundaryAgent acts unaided
✓Propose the next round from what the last one returned.
✓Schedule runs against instruments, reagents and the deck.
✓Test the preconditions, start an authorised run, and stop it on an alarm or a failed check.
✓Write the run record — what ran, on what, with which materials, under which plan version.
Write actions run only inside the approval boundaries agreed during implementation. Widening one is not one of them.
Releasing a result as a reportable value.
Restarting a run after a safety stop.
Widening the envelope or the condition ranges.
Starting an unattended run nobody has agreed to.
Example output
One proposed round, annotated
Everything the agent proposes is attached to the envelope and the preconditions it was read against.
Run proposal · single campaign roundIllustrative example
Campaign
Round
Instruments
Proposed
Confidence
Output status
Buffer-condition screen
3 of an agreed 5
Handler, reader, incubator
Next condition set and run order
86%
Run data, not reportable
As receivedThe campaign, the round it is on and the rig booked for it — nothing on this side is proposed by the agent.
Evidence usedEnvelope as approvedLast round's readoutsCalibration and lot status
Why these conditionsThe last round pushed at the top of an allowed range, so the next set steps back inside it.
ActionApproveAmendSend back
What the score decidesBelow the configured threshold the round waits for the scientist, not the run queue.
Value
Where AI adds value
The same four claims, placed at the point in the workflow where each one applies.
Where the value landsValue 01 – 04
Every round in a campaignFrom your scheduler and instruments
03Planning & scheduling
Test every run against its envelope
Use the approved envelope, the instrument and calibration state, the reagent lots and the rules the deck has to respect.
01Approved path
Turn a returned round into the next one
The last round's readouts, the envelope and the deck are read together, so a scientist opens a proposed round rather than an empty plan.
02Human review
Stop the run before the deck does
A failed precondition or an alarm stops the run, parks the instrument and calls the person named on it; a run that degrades with neither is not caught here.
04Build an evidence trail
Retain the envelope version, the preconditions read, the conditions proposed, the instruments and lots used, the stop and each scientist amendment — on both paths.
Integrations
Typical integrations
Five system groups connect to the same agent. Which of them are in scope is decided in discovery.
Integration availability depends on the client's existing systems and API access.
Agent controls
Six layers between the model and the deck
Each control wraps the one inside it. A proposed round clears every layer before an instrument is asked to move.
L6 · Outermost — last line of defenceInward → L1 · closest to the model
L6Trail and safe modeProvenance is written with the run, and automation narrows if evaluations or production signals degrade.Record
L5Watch and stopRuns are watched against expected behaviour and stopped on an alarm or a failing check; a failure that raises neither is not seen here.Stop
L4Human authorisationDefine what may start unattended and what waits for the lab head, the named scientist or a safety owner.Gate
L3Deck separationMaterials and samples your rules keep apart are held off one deck and one window; a conflict is raised here, not resolved.Separate
L2Precondition gateInstrument state, calibration status, reagent lots and the loaded deck are read before a step runs; a failed read stops the build.Verify
L1Envelope checkConditions are tested against the ranges the envelope names; one it does not describe is held, and an uncovered range is measured here.Test
Model coreRound proposed — conditions, run order, instrument and reagent bookings, and confidence
L1 – L2Decide whether a run may be built
L3Keeps two things off one deck
L4 – L5Gate the start; stop on an alarm or a failing check
L6Keeps the provenance, narrows automation
How Nestack evaluates it
Evaluate the whole campaign — not only the run that finished.
Coverage runs the whole depth of the workflow, and every layer is cut by slice.
Surface — the round a scientist is asked to approve
Depth of coverage ▼
E1Proposal evaluationWere the conditions inside the envelope and tied to the last round?
E2Precondition evaluationWere instrument, calibration and lot state read as they actually were?
E3Schedule evaluationDid the deck and the window keep incompatible work apart?
E4Stop evaluationDid the run stop when it should have, and park where it should?
E5Slice evaluationHow does run quality change across specific run cohorts?
E6Business outcomeHow many runs were aborted, repeated or thrown away?
Floor — the run record the lab works from
Failure modes
Where each failure originates in the agent
Seven failure modes plotted against the five stages of the agent lifecycle.
Agent lifecycleDirection of processing →
01 · Planning2 modes
AL-01
Condition proposed outside the envelope
A range the envelope never covered is treated as allowed.
AL-02
Optimiser chases an artefact
The readout climbs for a reason that is not the effect.
Stage proposesThe next round's conditions, inside the envelope
02 · Preconditions2 modes
AL-03
Run started on a lapsed calibration
Qualification expired between the check and the run.
AL-04
Deck does not match the plate map
Tips, plates or a lot were changed after the plan was built.
Stage checksInstrument state, calibration, lots and the deck
03 · Scheduling1 mode
AL-05
Incompatible work booked together
Two reagents or samples share a deck or a window they must not.
Stage booksInstruments, windows and what may share a deck
04 · Execution1 mode
AL-06
Overnight failure raises nothing
A run degrades with no alarm and no failing check until morning.
Stage runsOnly where the envelope and the authorisation allow
05 · Record / Version1 mode
AL-07
Manual intervention never recorded
Someone opens the enclosure and the record does not show it.
Stage recordsPlan version, materials, instruments and the stop
Sev-1 · a run puts material or a rig at riskSev-2 · the data cannot be traced backSev-3 · the round is wasted and repeated
Aborted, repeated and discarded runs, counted where a precondition, a deck or a stop was wrong and never where the science simply did not work. An overnight round has nobody in the building when the deck turns out wrong.
Slice performance — reported separately, not only in aggregateIllustrative example
Slice
Failure rate
Lift
Lift vs. threshold
Status
Unattended overnight runs
4.3%
2.8×
Review
Newly integrated instruments
3.6%
2.3×
Review
Runs after a lot change
2.5%
1.6×
Watch
Repeats of a known condition
1.4%
0.9×
Normal
Bar: lost-run lift vs. known-condition baseline · scale 0–4.0× · tick at the 2.0× threshold2 of 4 slices over threshold
Evidence-linked improvement
An aborted run says something about the plan
What stopped a run at night is worth more than the round it cost. The precondition, the deck rule or the range behind it is what has to move.
Improvement cycle · five stagesSwitchback — the path turns at Improve and returns at Learn
01Detect
Aborts, repeats and failed preconditions rise in one run cohort.
02Diagnose
The night's logs and the deck as found are set beside what the plan assumed.
03Improve
The rule that failed is rewritten, and the lab head signs it before a run uses it.
04Verify
The failing conditions are run again on a rig nothing else is relying on.
05Learn
What the night left behind is written into the checks the next run starts from.
Learn → DetectThe return edge. The envelope itself moves only with the lab head and the safety owners who set it — a limit the agent runs against is not a limit the agent may widen.
Typical build scope
Twelve workstreams across six weeks
The build scope read against the delivery timeline. Week structure follows the six-week plan — discovery, envelope and instruments, planning and scheduling, evaluation, integration, then production validation and handover.
WorkstreamWeek 1Week 2Week 3Week 4Week 5Week 6
01Workflow discovery and boundary definition.
02Envelope definition with the lab head and safety owners.
03Instrument, driver and scheduler assessment.
04Precondition reads — calibration, lots, deck.
05Campaign planning and condition proposal.
06Scheduling and deck-separation rules.
07Watch, stop and safe-state routing.
08Lab-head, scientist and safety-owner review workflow.
09Run record, provenance and version capture.
10Evaluation suite and regression runs.
11Instrument integration and rehearsed runs.
12Observability, deployment and Agent Care handover.
12 workstreams · 6 weeks · bar shows the weeks a workstream is active — several run in parallelFinal scope and sequence confirmed in discovery
Engagement tiers
What each tier includes
Rows are the capabilities named in each tier's scope. Higher tiers include everything below them.
Capability✓ in scope · — not at this tierPilotOne campaign, one rigProductionProduction integrationAdvancedSeveral rigs / sites
Introduced at Pilot
Envelope definition and checking✓✓✓
Precondition and calibration-status reads✓✓✓
Deck separation and incompatibility rules✓✓✓
Watch, stop and safe-state routing✓✓✓
Run record and provenance capture✓✓✓
Lab-head and safety-owner authorisation✓✓✓
Baseline evaluation✓✓✓
Introduced at Production
Scheduler and instrument integration—✓✓
Observability and evaluation—✓✓
Introduced at Advanced
Several rigs, sites and instrument families——✓
Multi-programme and enterprise controls——✓
Build priceFrom $5,000From $8,000Custom quote
Final build priceConfirmed after discovery based on the instruments and scheduler in scope, campaign types, run volume, authorisation controls, site coverage and deployment requirements.
Separate from buildBuild pricing is separate from recurring Agent Care, which covers managed monitoring, evaluations, incidents and verified improvements after launch.
What we need from you
What you bring, and what we build with it
Each input maps to a piece of build scope and a week in the delivery timeline.
You bringWe build with it
01The envelope you are willing to let a run happen inside→Envelope definition with the lab head and safety ownersWeek 1
02Where your approval boundary sits and who signs→Approval-boundary definition and authorisation rulesWeek 1
03Driver, scheduler and calibration-record access→Instrument, driver and scheduler assessmentWeek 2
04A campaign your scientists have already run by hand→Campaign planning, condition proposal and precondition readsWeek 2
05Your own rules for what may not share a deck→Scheduling and deck-separation rulesWeek 3
06Runs that went wrong — an abort, a bad deck, a lost night→Evaluation suite, regression runs and failure-mode testingWeek 4
07Named scientists, the lab head and your safety owners→Review workflow, then rehearsed runs and production validationWeeks 5–6
Nothing else is requiredDeployment, documentation and Agent Care handover are ours.
Delivery timeline
Four phases across six weeks
Phases are drawn over the weeks they actually occupy. Week 5 is the first week an instrument moves for the agent, and it is rehearsed before it is live.
PhaseW1W2W3W4W5W6
DiscoveryW1
BuildW2 – W3
EvaluateW4 – W5
Pilot & LaunchW5 – W6
Week focusW1Envelope and the stop agreed with your safety ownersW2Instruments, drivers, scheduler and precondition readsW3Condition proposal, scheduling and deck separationW4Run-record capture, stop tests and cohort slicesW5Instrument links, rehearsed runs and amendmentsW6Production validation, live runs and Agent Care handover
Reading the bandBars sit over named work, and an empty week in a row is empty on purpose. The envelope and the stop are built in week 1; no instrument moves under a rule that was not agreed there.
At the end of W6Your scientists have watched runs the agent planned, including at least one they stopped. Envelope checks, precondition reads and the version map move to Agent Care.
DurationSix-week plan shown · typical delivery 4–6 weeks depending on scope confirmed in discovery.
Next step · Biotechnology AI agent
Build an experimentation agent around the rigs you already run.
Show us one campaign, the instruments it runs on and what your lab does today when a run fails overnight. The stop matters more than the envelope, and who it wakes matters more than the stop. We design all three in week 1 with your lab head and your safety owners, and rehearse them before an instrument moves.
