Eight agents, one factory conductor

Each agent owns a physical step, a set of tools and an autonomy level. The factory orchestrator owns the lot, the handoffs and the human approvals.

Book a line assessment Read the tool reference

Edge + cloud
24 tool families
Approval gates on validated parameters

Who hands off to whom

Select any node to inspect the owning agent, the tool it invokes, its inputs and outputs, and its status in RUN-4417.

RUN-4417 · Velo Trainer 2 · Line B · plant HCMC-03agent graph

View as table

Text equivalent — nodes, owning agent, dependency and status
#	Node	Agent	Depends on	Status
01	Twin simulate	Yield-and-Energy	—	SUCCEEDED
02	Cut and nest	Cut-and-Nest	Twin simulate	SUCCEEDED
03	Stitch upper	Stitch-and-Upper	Cut and nest	SUCCEEDED
04	Mould sole	Mold-and-Sole	Stitch upper	SUCCEEDED
05	Last upper	Last-and-Bond	Mould sole	SUCCEEDED
06	Bond sole	Last-and-Bond	Last upper	APPROVAL
07	Inspect pair	Quality-and-Fit	Bond sole	SUCCEEDED
08	Finish and pack	Robot-and-Assembly	Inspect pair	SUCCEEDED

The agents in full

What each one perceives, what it controls, and how it is measured.

Cut-and-Nest

Reads every hide and roll, marks scars, brand marks and grain direction, then solves a defect-aware nest across the live size curve and drives the blade or laser cutter.

Leather yield +8.3 pts
Scrap −40%
Nest solve under 15 s

Stitch-and-Upper

Controls stitch tension, seam alignment and upper assembly sequence across stitching stations, catching skips, puckering and edge deviation before the upper moves on.

Seam defects −57%
Tension held ±4%
14 stations per line

Mold-and-Sole

Runs compression and injection moulding for EVA, PU and rubber — temperature, pressure, dwell and cure — to hold midsole density and outsole geometry against material-lot drift.

Density variance −62%
Short shots ≈ 0
Cycle time −9%

Last-and-Bond

The make-or-break agent: lasting fit under heat and tension, then primer, adhesive, activation temperature, press pressure and cure — the parameters that decide whether a shoe lasts years or delaminates in months.

Delamination returns −71%
Bond-gap escapes 0
Window held in 96% of shifts

Quality-and-Fit

Vision and metrology on bond line, seam, sole geometry, cosmetics and 3D fit — predicting defect and return risk per pair, not per sampled batch.

First-pass yield 98.2%
Escape rate 0.14%
100% pairs inspected

Robot-and-Assembly

Drives robotic material handling, upper and sole assembly, cleaning, lacing, insole insertion and packing on Jetson-class edge units.

Manual handling −38%
Pack accuracy 99.9%
6 cell types supported

Yield-and-Energy

Optimises material waste, energy per pair, line balance and changeover scheduling — and owns the shoe-and-line digital twin that simulates a run before it starts.

Energy per pair −18%
Changeover −46%
Twin fidelity ±2%

Footwear-Knowledge

Retrieval over patterns, BOMs, construction specs, material and adhesive datasheets, and quality procedures — so every agent decision cites the clause or datasheet behind it.

Citations on 100% outputs
Multi-tenant isolated
Versioned per factory

A lot passing between agents

The handoff is the risky moment in any multi-agent system. Footeon makes each one an explicit, logged transfer with a validated artefact.

01 twin.simulate_run SUCCEEDED 2m 40s

Simulated the whole lot in the shoe-and-line twin before a single hide was cut — nest layout, mould recipe and bonding process window for EVA midsole and rubber outsole at 71% relative humidity.

inputs · outputs
in: pattern velo-trainer-2, hide batch LH-2291, mould set M-14 · out: nest v7, recipe R-08, bond window 62–68 °C
02 cut.nest_optimize SUCCEEDED 4m 12s

Scanned 64 hides, marked 1,842 scars and brand marks, then solved a defect-aware nest across the EU 36–46 size curve. Leather yield came in at 87.4% against a 79.1% plant baseline.

inputs · outputs
in: 64 hides, 1,842 defects · out: yield 87.4%, 12.3 s solve, 22 nested part groups
03 stitch.tension_control SUCCEEDED 11m 08s

Held stitch tension inside ±4% across 14 stitching stations, caught three skipped stitches on the lateral overlay and re-ran those uppers before they reached lasting.

inputs · outputs
in: seam map v3, thread lot TL-77 · out: 3 skips corrected, 0 escapes
04 mold.density_control SUCCEEDED 9m 51s

Trimmed press temperature and dwell on P-3 as the EVA lot ran slightly high in blowing agent, holding midsole density at 0.24 g/cm³ within ±1.8% and cure at 214 seconds.

inputs · outputs
in: EVA lot EV-1180, recipe R-08 · out: density 0.24 g/cm³, cure 214 s
05 last.fit_control SUCCEEDED 6m 33s

Lasted the upper over last set L-42 under a heat and tension profile tuned to this leather lot. Toe-lasting deviation held at 0.6 mm; heel seat measured true on every pair sampled.

inputs · outputs
in: last set L-42, 78 °C, profile T-2 · out: deviation 0.6 mm
06 bond.window_control APPROVAL 3m 04s

Humidity climbed to 71% mid-shift, pushing the primer flash-off outside the validated window. The agent proposed +2 °C activation and +0.3 bar press pressure, and paused for the quality engineer to approve — the change touches a validated bonding parameter.

inputs · outputs
in: primer PR-9, adhesive AD-22, RH 71% · out: proposed activation 66 °C, awaiting sign-off
07 quality.inspect SUCCEEDED 5m 47s

Bond-line vision, 3D fit metrology and cosmetic scan on every pair. First-pass yield 98.2%; 21 pairs routed to rework for adhesive squeeze-out, none for bond gaps.

inputs · outputs
in: 1,200 pairs · out: FPY 98.2%, 21 rework, 0 bond-gap escapes
08 robot.finish_pack SUCCEEDED 18m 22s

Robotic cleaning, lacing, insole insertion and carton packing to spec C-4, with per-pair traceability written back to the MES and the brand’s quality portal.

inputs · outputs
in: carton spec C-4 · out: 1,179 pairs packed, 21 held

Handoff contract

Each transfer carries a typed artefact — a nest, an upper record, a sole record, a bonded pair — validated against schema before the next agent accepts it.

Rejected artefacts stop the line, not the next station
Every handoff is timestamped and attributable
Feedback edges let Quality re-open Cut on systemic defects

How an agent argues its case

The bond agent’s trace from RUN-4417 — plan, observation, thought, action, outcome.

Plan Hold the validated bond window across the shift

Adhesive datasheet AD-22 §4.2, construction spec VT2-CEM-03, target activation 62–68 °C.
Observation Humidity drifted from 58% to 71%

Station sensor line-b/rh-04, corroborated by the plant HVAC historian.
Thought Primer flash-off now exceeds the window at current line speed

Extended flash-off is the dominant precursor to edge delamination in cemented construction.
Action Request approval for activation 66 °C, press +0.3 bar

Validated parameter — agent is configured to act-with-approval and cannot proceed alone.
Outcome Approved, executed, verified

Bond-line vision confirmed zero gap escapes across 1,200 pairs; change retained in audit trail A-88213.

Why traces matter

Every Footeon decision that touches a cut, a mould, a last or a bond line is recorded as plan, thought, action and observation — with the material lot, sensor reading and spec clause that justified it. Quality engineers can replay any pair.

Traces are retained for the full product-liability window and exported to your MES and quality system.

Agents, tools and default autonomy

Defaults are conservative. Plants raise them workflow by workflow.

Agent roster with tool families and default autonomy level
Agent	Tool family	Default autonomy	Escalates to
Cut-and-Nest	cut., nest.	Act	Cutting room lead
Stitch-and-Upper	stitch.*	Act	Line supervisor
Mold-and-Sole	mold., press.	Act with approval	Process engineer
Last-and-Bond	last., bond.	Act with approval	Quality engineer
Quality-and-Fit	quality.*	Act	Quality director
Robot-and-Assembly	robot.*	Act	Cell technician
Yield-and-Energy	yield., twin.	Recommend	Plant manager
Footwear-Knowledge	knowledge.*	Observe	—

Agents are configuration, not a rewrite

Scope an agent to a line, bind its tools, set autonomy and approval routing.

line_b_bond_agent.pypython

# Bind the Last-and-Bond agent to Line B with act-with-approval on
# validated bonding parameters.
from footeon import Agent, Autonomy, Approval

bond = Agent(
    name="last-and-bond/line-b",
    plant="HCMC-03",
    tools=["last.fit_control", "bond.window_control", "bond.check_window"],
    knowledge=["spec:VT2-CEM-03", "datasheet:AD-22", "datasheet:PR-9"],
)

bond.set_autonomy(
    default=Autonomy.ACT,
    validated_params=Autonomy.ACT_WITH_APPROVAL,   # activation temp, press, cure
)

bond.on_approval(
    Approval(role="quality_engineer", sla_minutes=15,
             attach=["sensor_evidence", "datasheet_clause", "twin_prediction"]),
)

bond.deploy(edge="line-b/jetson-01", shadow_days=14)

Deployment defaults

Fourteen days of shadow mode before any write
Signed model artefacts, one-command rollback
Approval SLA breach falls back to last known-good recipe
All tool calls schema-validated at the edge

Developer hub →

Agents that remember your factory

The scarce craft in footwear lives in the hands of pattern cutters, lasting operators and sole technicians. Footeon captures it, per factory, without leaking it.

Quality-and-yield history

Every run, defect, rework and return, indexed by model, material lot and station.

Craft memory

Per-cutter and per-technician performance patterns that encode what actually works on this line.

Retrieval with citations

Patterns, BOMs, construction specs, material and adhesive datasheets, permission-aware and cited.

Versioned and scoped

Memory is versioned per tenant and entity — it improves over time and never crosses customers.

Synthetic augmentation

Rare delamination, hide-scar and fit-fault variations generated rather than waited for.

Continuous evaluation

Golden datasets and LLM-as-judge gate every model and prompt change in CI.

How the agents are graded

Agents are measured on plant outcomes, not model benchmarks.

−18% energy per pair across cutting, moulding and lasting
0.14% defect escape rate to the brand’s QC gate
24 agent tool families shipped across the platform
99.95% factory-edge runtime availability target

What keeps an agent inside the lines

Guardrails, approvals, sandboxing and an audit trail — the four things that make autonomy acceptable in a plant.

Certifications and posture

In progress SOC 2 Type II — audit window open [ASPIRATIONAL]
In progress ISO 27001 — controls implemented, certification pending [ASPIRATIONAL]
Supported GDPR — EU data residency and DPA available
Supported Per-tenant isolation for patterns, lasts and construction IP
Supported On-premises deployment with zero telemetry egress

Agent-specific assurances

Graduated autonomy: observe → recommend → act-with-approval → act, set per agent and per parameter
Human-in-the-loop approval gates on every validated bonding, lasting and moulding parameter
Immutable, quality-grade audit log of every agent action, approval and setpoint write
Guardrails and schema validation on tool calls; fail-closed to last known-good recipe
Sandboxed factory-edge runtime with signed model artefacts and version rollback
Scoped permissions per line, per station and per operator role

Read the security overview →

Agent questions

It stops and asks. Every agent runs at a configured autonomy level: observe, recommend, act-with-approval, or act. Bonding, lasting and any parameter under a brand or safety validation defaults to act-with-approval, and every approval is written to an immutable audit log with the sensor evidence that triggered it.

Put one agent on one line

Fourteen days of shadow mode, then graduated autonomy. Start with the step that costs you most.

Book a line assessment See pricing