Nine agents that each own a process step and a hard metric
Treadum agents are not chat assistants bolted onto a dashboard. Each one perceives its station, proposes bounded actions, and is measured against scrap, first-pass yield, uniformity, cure energy or takt — under a single factory orchestrator.
The loop runs across every process step
What each agent owns
Every agent has a scope, a set of signals it must be able to see, and a metric it is accountable for.
Mix-and-Compound
Banbury and downstream mixing control — batch rheology, dispersion, temperature and energy per kilogram, tuned against the compound spec instead of a fixed recipe clock.
Extrude-and-Calender
Tread and sidewall extrusion plus fabric and steel-cord calendering — gauge, width, profile and tension held to spec with closed-loop die and roll adjustment.
Build-and-Splice
Green-tire building — component placement, splice overlap and open-splice risk, bead seating and drum tension, verified per component instead of per audit sample.
Cure-and-Mold
Press and mold control — cure-state estimation per tire, adaptive dwell, mold allocation and preventable over-cure energy removed from the cycle.
Defect-and-Inspect
Vision, X-ray and shearography fusion for bubbles, separations, bare spots, cord shift and splice defects, with a hold decision made at the station.
Uniformity-and-Balance
Radial and lateral force variation, conicity and balance predicted upstream and graded downstream, so rejects are prevented rather than sorted.
Yield-and-Takt
Scrap recovery, WIP movement, line balancing and takt across mixing, building and curing so the constraint moves where you want it.
Robot-and-Handling
Robots, AGVs, green-tire transfer and press load/unload sequenced under one motion plan with bounded, envelope-checked actions.
Quality-and-Conformance
Right-first-time, non-conformance workflow, tire genealogy and traceability wired to IATF 16949, DOT and UNECE evidence needs.
How nine agents avoid fighting each other
Local optimization is how plants get a great mixing line and a starved press bank. The orchestrator arbitrates.
- Objectives: The site sets weighted objectives — first-pass yield, uniformity grade mix, cure energy, takt and due dates — versioned like any other config.
- Proposals: Each agent proposes actions with an expected effect on its metric and a confidence, plus the evidence behind it.
- Arbitration: The orchestrator scores proposals against the shared objective function and the twin, rejecting locally good but globally costly moves.
- Commit: Winning actions execute inside the envelope, or escalate to the right approver with the trade-off made explicit.
- Attribution: Outcomes are attributed back to the agents that caused them, so credit and blame are measurable, not rhetorical.
Four agents, up close
The wedge agents most plants start with.
Every splice measured, not sampled
Reads drum servos, component feed, laser profile and station vision to score overlap, gap and open-splice risk per component, then corrects tension and placement inside the envelope before the green tire moves on.
- Signals: drum encoders, servo torque, laser profile, station vision
- Metric: splice defect rate and first-pass yield at building
- Actions: drum tension, placement offset, hold
Cure state per tire, not per recipe
Estimates cure state from press temperature and pressure traces, mold history and the compound batch that actually went in, then adapts dwell within signed limits and flags molds drifting toward rework.
- Signals: press traces, mold ID and wear history, batch rheology
- Metric: cure energy per tire and under- or over-cure escapes
- Actions: dwell adjustment, mold allocation, hold
One finding, localized to a station
Fuses X-ray, shearography, thermal and camera streams on a shared timeline, classifies the defect family, and traces it back to the component, builder, drum, batch and press that produced it.
- Signals: X-ray, shearography, line-scan vision, thermal
- Metric: escape rate and false-hold rate
- Actions: hold, route to review, raise station flag
Predict the grade before the machine says it
Learns the mapping from build and cure evidence to radial and lateral force variation, conicity and balance, so uniformity rejects become an upstream correction instead of a downstream sort.
- Signals: build placement, splice geometry, cure trace, grading results
- Metric: grade A rate and uniformity reject rate
- Actions: upstream correction proposals, grade prediction, hold
Model strategy
A router puts the best and cheapest model on each step; high-volume steps move to fine-tuned open models to control cost of goods.
Fine-tuned defect sensing
Bubble, separation, splice, cord and bare-spot models trained per site on linked inspection and genealogy data, with X-ray and shearography fusion at the station.
Rheology, build and cure models
Compound-rheology, build and splice, and cure-state models run at the factory edge, with time-series prognostics for mixer and press equipment health.
Frontier models where they earn it
Tire-process and safety reasoning and engineer question answering use frontier models; the high-volume perception path does not.
pgvector over your specs
Defect, splice and X-ray imagery plus recipe and construction documents retrieved with permission-aware, tenant-isolated search and enforced citations.
Per-plant, per-engineer
Yield and quality history plus builder and engineer performance memory captures the craft that is walking out the door, scoped per tenant.
Agent, metric, gate
Each agent graduates its own autonomy gate independently. A site can run bounded curing and assist-only building.
| Agent | Primary metric | Typical entry gate | Escalates to |
|---|---|---|---|
| Mix-and-Compound | Batch-to-batch rheology variance | Assist | Process engineer |
| Extrude-and-Calender | Gauge deviation and scrap length | Bounded | Process engineer |
| Build-and-Splice | Splice defect rate | Assist | Building supervisor |
| Cure-and-Mold | Cure energy and cure escapes | Bounded | Process engineer |
| Defect-and-Inspect | Escape rate and false holds | Bounded | Quality engineer |
| Uniformity-and-Balance | Grade A rate | Assist | Quality engineer |
| Yield-and-Takt | Takt attainment and scrap | Advisory | Plant director |
| Robot-and-Handling | Handling cycle time and incidents | Assist | Automation lead |
| Quality-and-Conformance | Right-first-time and NCR closure | Advisory | Quality manager |
The engineer stays the authority
Disagreement is the most valuable signal in the plant
When an engineer rejects or modifies a proposal, Treadum captures the reason as structured data, not a free-text note nobody reads. Those overrides are the fastest path to a site model that matches how this plant actually runs.
- Structured override reasons per action class
- Reviewed weekly with the process team
- Fed into the next gated site release
{
"action": "cure.dwell",
"proposed_delta_s": -9,
"committed_delta_s": -4,
"decision": "modified",
"reason_code": "mold_wear_unmodeled",
"engineer": "r.menon",
"evidence": ["mold-41-wear-log", "cure-trace-8841236"],
"training_use": true
}How teams describe working with the agents
“We had three people watching X-ray images and still shipped uniformity rejects. Treadum flagged the cord shift at the building drum, not four hours later at final inspection.”
“The cure agent found 40 seconds of margin on a construction we had run the same way for eleven years. It proved it in shadow mode before it touched a press.”
“Genealogy is the part I did not expect to care about. Every tire now has a linked record from batch to grade, and audit prep went from weeks to an afternoon.”
[PLACEHOLDER] Design-partner quotes are illustrative until pilot references are published.
What a healthy agent fleet looks like
Common questions about autonomy
Yes. Agents are independently licensed, deployed and gated. Most plants start with one — usually defect-and-inspect or cure-and-mold — and add neighbours once the wedge metric moves.
No. They publish proposals and evidence to the orchestrator, which arbitrates against the site objective function. That keeps behaviour auditable and prevents feedback loops between stations.
The objective function is weighted and versioned by the site, and the twin evaluates the trade-off before commit. An action that improves takt at the cost of predicted grade mix is rejected or escalated.
Site-pinned versions with gated releases. A candidate release must clear golden datasets, including rare defect families, before it can be promoted, and any release can be rolled back without touching the schedule.
Put one cell on autonomy in 90 days
Pick one wedge — splice inspection, X-ray defect detection, cure-state optimization or uniformity prediction. We baseline it, run shadow mode, then graduate to bounded autonomy under a signed control envelope.
Pilots start in shadow mode. No line changes until accuracy and safety gates pass.