Declared goals, typed steps
Every run starts from a declared goal, is planned as a graph of typed steps and ends in a written outcome. Nothing is implicit.
Features
Everything the plant needs to trust a machine that acts.
Autonomy on a glass line is earned with evidence, limits and rehearsal. These are the features that provide all three.
Core
The features that make autonomy safe
The platform is built so a plant can trust software that moves a furnace. Each feature below exists because something would go wrong without it.
Checked before every write
Autonomy level, tag allow-list, rate and magnitude limits, interlock state and operator presence are evaluated before a setpoint moves.
Named people, logged decisions
Anything above the site threshold waits for a glass technologist. The request, the reasoning and the decision are all in the log.
Append-only, hash-chained
Every tool call, argument, result and reasoning step, retained on the plant's storage and exportable for a quality audit.
Every piece, traced
Batch, melt, forming, lehr curve, defect map, stress map, agent actions and human approvals, written per lot and per ware.
Simulate first
Candidate recipes ranked on Glastwin before a move is proposed; approved moves replayed before promotion to higher autonomy.
For technologists, not data scientists
Approve, correct and comment on recommendations. Corrections train the models; the console is where the craft gets captured.
Measure before you trust
Every pilot starts observing. The baseline is measured by the plant's own genealogy before a recommendation is scored.
Best or cheapest per step
High-volume perception and control on fine-tuned open models; glass-technology reasoning on frontier models; every output cited.
Run timeline · scenario
One scenario run, step by step
Every Glasent run is an ordered, inspectable sequence. This is scenario_fl2_01 on FL-2 · float line · clear soda-lime: Thickness change FL-2 · 6 mm to 4 mm clear float, residual stress inside spec, zero escaped seeds. It is a worked scenario that shows the shape of a run, not a measured customer result.
scenario_fl2_01
FL-2 · float line · clear soda-lime
scenario
SUCCEEDED
-
01
ingest.order
SUCCEEDED
0.9 s
Plant Orchestrator · Pulled the 4 mm clear float spec, optical-grade tolerances and the standing energy window from plant MES; locked the target envelope for the run.
-
02
twin.simulate
SUCCEEDED
41 s
Glastwin · Simulated 36 candidate transition recipes across furnace pull, tin-bath ribbon speed and lehr curve; ranked them on seed risk, residual stress and energy per tonne.
-
03
melt.pull
SUCCEEDED
6 m 20 s
Meltrix · Stepped furnace pull toward the new ribbon mass flow while holding melt temperature and fining; chemistry stayed inside the composition window.
-
04
form.ribbon
SUCCEEDED
4 m 05 s
Formeon · Raised ribbon speed and re-angled the top rollers to thin the ribbon toward 4 mm; thickness converged inside the design tolerance.
-
05
anneal.curve
SUCCEEDED
3 m 12 s
Anneon · Re-shaped the lehr cooling curve for the thinner, faster ribbon so residual stress stays inside spec at the higher speed.
-
06
inspect.ribbon
SUCCEEDED
continuous
Seedscan · 8 camera, optical and stress stations streaming; a seed cluster flagged at the ribbon edge and attributed to the pull transient, routed to cullet.
-
07
yield.balance
SUCCEEDED
2.4 s
Plant Orchestrator · Re-sequenced cut sizes so transition ribbon routes to cullet recovery and good ribbon to the highest-value open order.
-
08
approve.human
APPROVAL
48 s
Plant Orchestrator · The second pull step exceeded the site autonomy threshold. Held for the glass technologist on shift; approved and written to the audit log.
-
09
handle.stack
SUCCEEDED
1 m 10 s
Panebot · Re-planned pick and stack paths for the thinner panes; plates flagged by Seedscan diverted to cullet, good plates stacked to rack A3.
-
10
ware.qualify
SUCCEEDED
1 m 02 s
Plant Orchestrator · Lot released with full genealogy: batch, melt, forming, lehr curve, defect map, stress map and the technologist's approval.
Tool calls · scenario
The agent shows its work
Every tool invocation, argument and result is written to an immutable, human-readable log, and every reasoning step is expandable. Nothing about a run is hidden from the plant.
- 02:14:03mes.read_orders({ line: "FL-2", next: "CLR-4MM" })
- 02:14:03ok spec locked · 4.0 mm · optical grade · stress limit set
- 02:14:04twin.simulate({ candidates: 36, horizon: "22 min" })
- 02:14:45ok best recipe #19 · seed risk low · stress inside spec
- 02:14:46policy.evaluate({ recipe: 19, autonomy: "L3" })
- 02:14:46ok 11 writes permitted · 1 write requires human approval
- 02:14:47scada.write_setpoint({ tag: "F1.PULL_SP", step: 1 })
- 02:21:07ok pull step 1 settled · fining stable · chemistry in window
- 02:21:08float.write_ribbon({ speed: "+", rollers: "re-angle" })
- 02:25:13ok thickness converging to 4.0 mm · flatness in tolerance
- 02:25:14lehr.write_curve({ zones: 12, profile: "4mm-fast" })
- 02:28:26ok residual stress inside spec at new belt speed
- 02:28:27vision.stream({ stations: 8, model: "seedscan-v3" })
- 02:28:41seed cluster · ribbon edge · attributed to pull transient
- 02:28:42cuopt.sequence({ objective: "cullet+energy" })
- 02:28:44ok transition ribbon routed to cullet · good ribbon to order
- 02:28:45approval.request({ action: "pull_step_2" })
- 02:29:33ok approved by the technologist on shift · logged
- 02:31:02isaac.plan_path({ stack: "A3", thickness: "4.0 mm" })
- 02:32:12ok paths validated in Isaac Sim · flagged plates diverted
- 02:33:15mes.log_lot({ lot: "FL2-4MM", genealogy: "full" })
- 02:34:17ok scenario run complete · stress in spec · 0 escaped seeds
Reasoning trace
Plan, thought, action, observation, exactly as the orchestrator would record it.
GoalMove FL-2 from 6 mm to 4 mm clear float with residual stress i…
Move FL-2 from 6 mm to 4 mm clear float with residual stress inside spec and zero escaped seeds.
PlanSimulate the transition on the twin first. Lead with furnace p…
Simulate the transition on the twin first. Lead with furnace pull so the ribbon mass flow is right before the forming change, then thin the ribbon, then re-shape the lehr curve. Watch the ribbon continuously; hold the second pull step for a human.
ThoughtThe standing recipe changes pull and ribbon speed together. On…
The standing recipe changes pull and ribbon speed together. On the twin that produces a fining transient that seeds the ribbon edge. Sequencing the pull step ahead of the forming change and holding fining removes most of it.
Actiontwin.simulate(candidates=36) returned recipe #19: pull in two …
twin.simulate(candidates=36) returned recipe #19: pull in two steps, forming change after fining settles, lehr curve re-shaped before belt speed rises.
ObservationRecipe #19 scored lowest on seed risk with stress inside spec.…
Recipe #19 scored lowest on seed risk with stress inside spec. Two candidates scored lower on energy but pushed residual stress over the site limit and were discarded.
ActionExecute recipe #19 under autonomy level L3: eleven setpoint wr…
Execute recipe #19 under autonomy level L3: eleven setpoint writes permitted, the second pull step routed to the glass technologist.
ObservationSeed cluster at the ribbon edge at 02:28:41, attributed to the…
Seed cluster at the ribbon edge at 02:28:41, attributed to the pull transient. cuOpt routed that ribbon to cullet recovery; no flagged plate reached a customer stack.
OutcomeScenario run complete. Thickness at 4.0 mm, residual stress in…
Scenario run complete. Thickness at 4.0 mm, residual stress inside spec, one approval gate, full genealogy written to the lot record.
In depth
Three features worth the detail
Latency, evidence and rehearsal are where a glass plant's trust is won or lost.
Latency
Perception at line speed
Container lines form thousands of pieces an hour and a float ribbon never stops. Seedscan runs on Jetson Orin at the station, with DeepStream and Holoscan ingesting synchronised camera, optical, thermal and stress streams and TensorRT models classifying seeds, stones, cords, checks and stress indicators.
- Sub-100 ms defect alerts on high-speed container lines · design target
- Sub-500 ms fused stress and flatness decisions on flat glass · design target
- 30 to 60 frames a second per station · design target
agent.defect_inspect RUNNING
| Stage | Where | Budget |
|---|---|---|
| Capture and sync | Station | Holoscan |
| Classify | Jetson Orin | TensorRT |
| Attribute and signal | Line edge | < 100 ms total · target |
Evidence
Every claim traces to a record
A recommendation names the signals it read, the recipe or standard it relied on and the limits it applied. A write names the approver. A pilot report reproduces every figure from the plant's own ware genealogy, not from our telemetry.
- Citations on every knowledge-layer answer
- Hash-chained log entries, exportable as JSON or CSV
- Baseline agreed in writing before the pilot starts
audit.write SUCCEEDED
| Field | Content |
|---|---|
| request | agent, tool, arguments, timestamp |
| reasoning | plan, thought, cited sources |
| policy | level, limits applied, verdict |
| decision | approver, time, outcome |
Rehearsal
The twin takes the risk first
Furnace campaigns are long and product qualifications are unforgiving, so the expensive mistakes belong in simulation. Glastwin ranks candidate recipes on seed risk, residual stress, energy and takt, and discards anything over the site limit before the orchestrator sees it.
- 10 to 100 candidates per grade or thickness change
- Discard rules on stress and seed risk are site policy, not model opinion
- Approved moves replayed in the twin before promotion to L3 or L4
agent.twin SUCCEEDED
| Criterion | Rule |
|---|---|
| Residual stress | Must stay inside spec |
| Seed risk | Must stay below site limit |
| Energy per tonne | Minimised among survivors |
| Takt | Inside the order window |
Multi-agent handoff
Agents negotiate, they do not collide
Two agents will want the same actuator. The orchestrator arbitrates on the run goal, not on who asked first, and the handoff is logged like any other step.
Contested move: furnace pull rate
-
01form_shape.request(pull hold) SUCCEEDED0.2 s
Formeon wants pull held while the ribbon thins, to protect thickness convergence.
-
02batch_melt.request(pull step) SUCCEEDED0.2 s
Meltrix wants the second pull step now, to settle fining before the seed rate climbs.
-
03orchestrator.arbitrate SUCCEEDED0.4 s
Seed risk outranks a short thickness excursion under the run goal "zero escaped seeds". Meltrix wins the actuator, and because the step is above threshold it goes to the technologist.
-
04form_shape.handoff(returned) SUCCEEDED90 s
Actuator returned; Formeon recovers thickness with roller angle instead. Both requests, the score and the reason are in the run record.
Arbitration rules
- Run goal first. Every request is scored against the declared goal, not the requesting agent's local objective.
- Safety and escaped defects outrank throughput. A stress fault in the field costs a recall; a thickness excursion costs minutes of ribbon.
- Time-boxed ownership. An agent holds a contested actuator for a bounded window, then must re-justify.
- Everything is logged. The losing request, the score and the reason all appear in the run record.
- Deadlocks escalate to a person rather than resolving by timeout.
Autonomy
Four levels, set per agent and per tag
A plant does not go from manual to unattended in one step. Glasent makes the level explicit, auditable and reversible at any time, and the first release plan is shadow, then assist, then graduated autonomy.
| Level | What the agent does | What the person does | When |
|---|---|---|---|
| L1 · Shadow and advisory | Observes, predicts and recommends setpoints with its reasoning | Enters every change manually; a baseline is measured | Pilot weeks 1 to 3 |
| L2 · Assist | Proposes a write; it executes on approval | Approves each write in the review console | Pilot weeks 4 to 8 |
| L3 · Bounded | Writes inside tag, rate and magnitude limits on low-risk loops | Approves pull steps, grade releases and anything above threshold | Pilot week 9 onward |
| L4 · Unattended | Runs the approved envelope without prompting | Sets the envelope; reviews the shift record | Planned, after graduated autonomy proves out |
How it works
A goal becomes a plan becomes a run
Nothing about a Glasent run is implicit. The goal is declared, the plan is simulated on the twin, each step is typed, and the outcome is written to the ware record.
The orchestrator plans, the agents act
A run starts as a goal, becomes a plan, and executes as a graph of typed steps. The active path is always visible, always logged, always reversible.
RUNNING APPROVAL SUCCEEDED
Autonomy is a dial, not a switch
Shadow, assist, bounded, then unattended inside an envelope. Each site sets the level per agent, per tag, per shift, and every write above the threshold waits for a named glass technologist.
Measured
Design targets, stated as targets
These are the numbers the architecture is built to hit and the pilot is built to measure. None is a customer result yet; every pilot report reproduces its figures from the plant's own ware genealogy.
90–120 days
Paid line pilot, shadow to assist to bounded writes
3–5
Design-partner plants in the first cohort
50k–250k
Synthetic rare-defect scenes per glass family · planned
1,500°C
Where the process starts, in furnaces that run for a decade
Developers
Define an agent, bound it, run it
The Glasent SDK is typed Python. Tools are declared with schemas and limits; the policy engine enforces them at call time, not in a review meeting.
What you get
- Typed tool definitions with unit-aware ranges and rate limits
- Deterministic replay of any historical run against a new model
- Local twin harness so a recipe is simulated before it is shipped
- Autonomy policy as code, versioned and reviewed like any other change
# Bound the anneal agent to twelve lehr zones on FL-2.
from glasent import Agent, Tool, Limit, Autonomy
lehr = Tool(
name="lehr.write_curve",
tags=["FL2.LEHR.Z01..Z12.TEMP_SP"],
limits=[Limit(max_step="4 C", per="60s")],
)
anneal = Agent(
id="agent.anneal_stress",
goal="residual stress inside spec, min energy",
tools=[lehr, Tool("optic.read_birefringence", read_only=True)],
# bounded writes; a technologist still gates pull steps
autonomy=Autonomy.L3,
# simulate on Glastwin before every write
verify="twin",
)
run = anneal.start(line="FL-2", product="CLR-4MM")
for step in run.stream():
print(step.name, step.status, step.duration)
CLI
Start a run from anywhere
The same run engine, the same policy checks, the same audit trail, from the terminal, the review console or the SDK.
$ glasent run "thickness change FL-2 to 4 mm" --autonomy L3
→ plan composed 10 steps · 1 approval gate
→ twin.simulate 36 candidates · best #19 · stress in spec
→ policy.evaluate 11 writes permitted · 1 held for human
→ executing melt.pull ....... ok 6m20s
→ executing form.ribbon ..... ok 4m05s
→ executing anneal.curve .... ok 3m12s
! seed cluster ribbon edge · attributed to pull transient
! approval required pull step 2 · above site threshold
→ approved glass technologist on shift · 02:29:33
→ run complete scenario · stress in spec · 0 escaped seeds
$ glasent runs show scenario_fl2_01 --format genealogy
From the plant floor
The problems we hear
No customer quotes yet; we are pre-launch. These are the three buyer personas the product is built for and the pain each one describes, in their own terms.
"The furnace has run for years on the same setpoints. The process never has. We find out a melt drifted when the cullet pile grows."
Plant / operations director · ICP persona
"I can tell you why a check appeared from the lehr curve and the gob weight. I cannot be at every line, and the people who could are retiring."
Glass technologist · ICP persona
"A missed seed is a reject. A missed stress fault is a pane that shatters in the field. I need genealogy on every piece, not a spot check."
Quality / reliability engineer · ICP persona
Integrations
It speaks plant, not cloud
Glasent reads and writes through the furnace, forming, lehr, inspection and MES systems already on the floor. No rip-and-replace, no parallel historian, no new HMI to learn.
Furnace SCADA and PLC, batch-plant weighing, redox and fining instruments
Setpoint reads and guarded writes over OPC UA and Modbus
Float-bath and IS-machine controls, gob-weight and timing systems
Gob, pull, ribbon speed and roller reads; guarded writes
Lehr zone controllers and belt drives
Zone temperature and curve reads; guarded writes
Camera, optical, thermal, polariscope and birefringence stations
Frames, stress maps, defect records, line-speed streams
Orders, grades, lots and ware genealogy
Spec and tolerance reads; genealogy and release writes
Time-series stores and lab information systems
Backfill, replay and lab chemistry
Robot cells, conveyors and stackers via NVIDIA Isaac
Pick, path and stack commands inside the safety envelope
SSO and RBAC via SAML or OIDC; NVIDIA Jetson Orin edge nodes
Named approvers, tag-level roles, sub-100 ms inference
Guardrails
An agent that can move a furnace needs a leash
Glasent writes to production equipment. Every capability is scoped, every write is policy-checked, and every action is written to an append-only audit log the plant owns.
- Bounded action space. Each agent can only write to an explicit tag allow-list, inside per-tag rate and magnitude limits.
- Policy engine before every write. Autonomy level, shift, product, interlock state and operator presence are all evaluated before a setpoint moves.
- Human-in-the-loop gates. Anything above the site threshold, pull steps, grade releases, safety-adjacent moves, waits for a named approver.
- Immutable audit log. Append-only, hash-chained, exportable, and retained on the plant's own storage.
- Hard fallback. Loss of the edge node, the network or the model returns control to the furnace, forming and lehr systems' last known-good state.
- Tenant and IP isolation. Compositions, forming recipes and defect libraries never cross a customer boundary. On-prem deployment available.
Compliance posture
| Standard | Scope | Status |
|---|---|---|
| SOC 2 Type I | Cloud control plane | RUNNING Planned in the first six months |
| SOC 2 Type II | Cloud control plane | QUEUED Planned in months six to twelve |
| IEC 62443 | Plant-edge OT security | RUNNING Design-aligned |
| ISO 9001 / IATF 16949 | Quality and genealogy records | SUCCEEDED Record formats supported |
| Container and safety-glass standards | Stress and defect conformance records | SUCCEEDED Record formats supported |
Enterprise
Standardise autonomy across the group
One policy model, one audit trail, one benchmark across every line in every plant, with the composition and forming models kept private to each site.
- SSO and role-based access down to the tag level
- Group-wide autonomy policy with per-site override and approval chains
- Cross-plant benchmarking on cullet, seeds, energy per tonne and first-pass quality
- VPC, on-prem and air-gapped plant-edge deployment options
- Custom composition, forming and tolerance models per site
- Dedicated deployment engineer and quarterly model review per plant
FAQ
Straight answers
The questions plant directors and glass technologists ask in the first meeting.
Yes, but only within an explicit tag allow-list with per-tag rate and magnitude limits, and only at the autonomy level your site has set. Every pilot starts in shadow mode, where Glasent predicts and recommends and a person enters everything. Writes come later, after the recommendations have earned it.
Control returns to your existing furnace, forming and lehr systems at their last known-good state. Glasent is a supervisory layer on top of the control system you already run, never a replacement for it, so an outage degrades the plant to its current way of running, not to a stop.
Shadow mode starts on the first day from existing SCADA, forming, lehr and inspection data. Defect and stress prediction improve as site history and labelled outcomes accumulate; the pilot plan sets a baseline period before any recommendation is scored.
Only if you choose cloud training. Compositions, forming recipes and defect libraries are tenant-isolated and never used to train another customer's models. On-prem training and an air-gapped plant edge are available for IP-sensitive producers.
You are, the same as with any control strategy, which is why every write is policy-checked, bounded, logged and reversible, and why anything above your risk threshold waits for a named approver. The audit log records the request, the reasoning, the limits applied and the human decision.
A 90 to 120 day line pilot in three stages: shadow mode to measure the baseline, assist mode where a technologist approves each recommendation, then bounded write-back on low-risk forming, annealing or inspection loops if the plant is satisfied with the results.
Get started
See these features on your line
Shadow mode on one line shows you the run record, the policy engine and the review console on your own data before any write is enabled.