A goal, tools and an autonomy level
An agent declares what it is for, what it may call and how far it may go without a human. Its runs are typed and logged.
The Chassira SDK is a design preview of how an agent is defined, bounded and run: tools with unit-aware ranges, rate limits and qualified envelopes, parameters that can never be declared writable, autonomy policy as code, deterministic replay of any historical shift against a new model, and a local twin harness. The API will change before release.
Illustrative scenario: Plant BIW-2 · EV crossover body shop to end of line · shift 1, one approval gate
Everything in the SDK is one of these.
An agent declares what it is for, what it may call and how far it may go without a human. Its runs are typed and logged.
Tools name tags, units and whether they are read-only. Write tools carry Limits and an envelope. Locked tags refuse to be declared writable.
max_step per period, absolute floors and ceilings, and the qualified envelope per joint family, colour block or program, enforced at call time by the policy engine and again at the OT boundary.
The level the plant has set for this agent, this tag, this shift. Approvals are routed to named humans above the threshold.
Start, stream, replay. Every step has an agent, a tool, a duration and a status, and the whole record is kept per body and vehicle.
Simulate a move on the plant twin before it is written. Required for any agent configured with it.
The Chassira SDK is typed Python. Tools are declared with schemas, limits and the qualified envelope, the policy engine enforces them at call time, not in a review meeting, and safety-critical parameters are locked for every agent. This is a design preview; the API will change before release.
# Design preview. Bound the paint agent to booth B-2; weld schedules stay with a human.
from chassira import Agent, Tool, Limit, Autonomy
booth = Tool(
name="booth.set_env",
tags=["B2.HUMIDITY_PCT", "B2.AIRFLOW_M3H"],
limits=[Limit(max_step="2 pt", per="10min", envelope="EV-CUV-dark-block")],
)
paint = Agent(
id="agent.paint_finish",
goal="film build in window, dirt and run risk minimised, zero repaint escapes",
tools=[booth, Tool("filmbuild.read", read_only=True)],
# bounded writes on booth environment
autonomy=Autonomy.L3,
# simulate on the plant twin before every write
verify="twin",
)
# torque specifications and test verdicts are locked for every agent
run = paint.start(plant="BIW-2", shift=1)
for step in run.stream():
print(step.name, step.status, step.duration)
The same run engine, the same policy checks, the same audit trail, from the terminal, the review console or the SDK. Design preview.
$ chassira run "shift 1 on BIW-2, program EV-CUV" --autonomy L2+
→ plan composed 10 steps · 1 approval gate
→ twin.simulate 18 candidates · best #6 · expulsion risk 0.06
→ policy.evaluate 6 writes permitted · weld_schedule held for human · torque_spec locked
→ executing stamp.form ....... ok continuous
! approval required weld_schedule W-14 current −2% hold +20 ms
→ approved body-shop launch engineer · 06:14:52
→ executing weld.join ........ ok continuous
→ executing frame.geometry ... ok continuous
→ executing paint.finish ..... ok continuous
! held pack PK-1182 · torque trace off curve · routed to rework
→ scenario complete escapes 0 · shift 1 released
$ chassira runs show run_41a9c2 --format as-built
Every Chassira run is an ordered, inspectable sequence. This is run_41a9c2 on Plant BIW-2 · EV crossover body shop to end of line · shift 1. Scenario completes in 10 steps with one approval gate. Weld schedule change on robot W-14 approved by the body-shop launch engineer, framing offset held inside the envelope, booth humidity trimmed before the predicted dirt defects appeared, one pack held at marriage on a torque trace, and the shift released with full as-built records.
Plant Orchestrator · Pulled the shift build sequence, program control plan and qualified envelopes from MES: 412 bodies, four trim variants, weld schedule set W-rev 7, framing targets per station, paint colour blocks, torque specifications. Locked every safety-critical parameter as read-only.
Vehicle-and-Line Twin · Simulated 18 candidate weld-sequence, framing-offset and booth-window combinations on the plant twin, ranked on predicted body geometry, expulsion risk at robot W-14, paint defect map and takt against launch cadence.
Stamp-and-Form · Blank feed on press line P-3 adjusted by 0.4 mm and lubricant dose raised 3 percent for the new coil lot as springback risk on the rear door inner climbed. Panel dimension held inside the window; genealogy handed to framing.
Plant Orchestrator · Candidate #6 lowers weld current 2 percent and adds 20 ms hold on robot W-14 to cut predicted expulsion. Weld schedules are crash-safety critical and stay with a human, so the change was held for the body-shop launch engineer and approved at 06:14:52.
Weld-and-Join · Approved schedule 7b applied on W-14. Current and force traces confirmed inside the window on the first 40 bodies; predicted expulsion risk fell from 0.31 to 0.06 and no geometry debt was introduced at the adjacent seams.
Frame-and-Geometry · Framing station FS-1 offset corrected by 0.3 mm on the left B-pillar datum inside the qualified envelope as in-line measurement trended toward the tolerance edge. Gap-and-flush vision confirmed the correction on the next 25 bodies.
Paint-and-Finish · Booth B-2 humidity trimmed 2 points and airflow rebalanced for the dark-colour block as the defect model predicted dirt on horizontal panels. Film build held inside the window on the first colour block.
Powertrain-and-Battery · Pack PK-1182 held at marriage after a torque-to-yield trace on fastener 7 fell outside the specification curve. Routed to rework with the trace attached; torque specification untouched, as always.
Quality-and-Anomaly · Gap-and-flush and paint-defect cameras, weld and torque traces and test results fused per body. Two paint dirt defects localised to booth B-2 zone 3 before the humidity trim took effect; both routed to spot repair, none escaped to final.
EOL-and-Calibrate · End of line proved alignment, ADAS calibration, leak, brake, roll and functional tests on every vehicle; two ADAS retests sequenced inside takt. Shift 1 released with as-built records, approvals and dispositions written for the quality lead.
Every tool invocation, argument and result is written to an append-only, human-readable log, and every reasoning step is expandable. Nothing about a run is hidden from the plant.
Plan, thought, action, observation, exactly as the orchestrator would record it.
Run shift 1 on plant BIW-2 inside the qualified envelopes, hold body geometry at the framing station, keep expulsion and paint defects from escaping, hold every torque to specification and prove every vehicle at end of line.
Simulate weld-sequence, framing and booth candidates on the twin first. Lead with stamping so panels arrive in dimension, then propose the weld schedule and hold it for a human, then correct framing inside the envelope, then set the booth ahead of the colour block. Fuse inspection continuously. Never touch a torque specification or a test verdict.
The standing weld schedule on W-14 runs current high enough to guarantee the joint on the thickest stack. With the new coil lot and panel dimension held at P-3, the twin predicts the same joint quality at 2 percent lower current with a 20 ms hold, and lower expulsion, because the panels arrive flatter.
twin.simulate(candidates=18) returns #6: blank feed +0.4 mm on P-3, weld current -2 percent and +20 ms hold on W-14, framing offset -0.3 mm at FS-1, booth B-2 humidity -2 points for the dark block.
Candidate #6 predicts geometry inside targets, expulsion risk 0.06, a lower dirt map and takt at 61 seconds. Three candidates scored better on takt but pushed predicted B-pillar variation past the tolerance edge and were discarded.
Execute #6: six bounded writes permitted on stamping, framing, booth, routing and sequencing; the weld-schedule write routed to the body-shop launch engineer; torque specifications and test limits locked.
Dirt precursor in booth B-2 zone 3 at 07:40:02 before the humidity trim took effect. Two defects localised and routed to spot repair; none escaped. One pack held at marriage on a torque trace at 08:05:19.
Scenario complete. Geometry inside targets, no expulsion escapes, paint defects caught in station, one pack reworked, two ADAS retests passed inside takt. Shift 1 released with full as-built records.
A car plant does not go from manual to unattended in one step. Chassira makes the level explicit, auditable and reversible at any time. The first release runs shadow mode, then assist mode, then bounded autonomy on low-risk loops. Weld schedules stay supervised until the joint family is qualified, and torque specifications and test verdicts are never written by an agent.
| Level | What the agent does | What the human does | Where it sits in a pilot |
|---|---|---|---|
| L1 · Shadow | Reads weld, metrology, vision, torque, test, robot and MES signals; predicts and recommends; writes nothing | Runs the line as today; compares predictions to audit and engineer outcomes | Design partners start here; the baseline is measured from the plant's own records |
| L2 · Assist | Proposes weld, framing, paint, routing, holding and sequencing moves with evidence, confidence, envelope and rollback path | Approves each action in the review console | Assist mode: accuracy and ROI proven against the signed baseline |
| L3 · Bounded | Executes pre-approved low-risk actions: booth environment, framing offsets, blank feed and lubrication, routing, holds, sequencing, inside PLC and safety limits and the qualified envelope | Approves weld-schedule changes, torque-related decisions and vehicle release | Graduated autonomy, after audited accuracy and ROI |
| L4 · Scaled | Supervised weld-schedule adaptation for qualified joint families, across lines and plants | Sets the envelope; reviews the shift record and evaluation results | Planned; after golden-dataset evaluation, twin validation and quality sign-off. Torque specifications and test verdicts never leave human hands |
A spot weld lasts a fraction of a second and a body passes a station every minute. Perception has to be local, deterministic and fast, so Chassira runs GPU inference at the plant edge and keeps training, simulation and optimisation in the cloud or on-prem.
8 to 32 cameras per process area at 15 to 60 FPS, plus torque, dimensional, thermal, robot and paint-instrument streams fused at the cell. Weld and gap alerts under 50 ms and paint or dimensional prompts under 150 ms are the design targets. Jetson Thor is planned for higher density.
Weld-defect, geometry-risk, paint-risk, torque and battery, and takt-risk models served at the plant edge with versioned rollback, routed by latency, tenant and cost, with canary rollout by line.
Weld, paint and gap vision, dimensional models, NeMo IATF and spec reasoning and GPU-accelerated stamping, springback and weld-geometry FEA surrogates. 150 to 600 GPU-hours per plant during onboarding, then monthly retraining from engineer corrections. Dedicated capacity is planned.
Body framing, weld sequence, robot reach, paint flow and line balance simulated before the shift. The pattern is production proven at BMW and Mercedes-Benz, per the architecture; Chassira's twin and OVX capacity are planned.
Rare weld defects, paint failures, geometry-drift modes, occluded joints and misbuild cases generated and validated against real distributions before promotion. 50,000 to 250,000 variants per program is the design target. Planned.
Line sequencing and JIS feeding, AGV routing, takt, line balancing, buffers and retest sequencing.
Robotic weld, paint, assembly and inspection cells and AGV and AMR line-feed autonomy, with skills validated in simulation before a cell runs them.
Telemetry ETL across weld and press parameters, measurements, paint readings, torque traces, battery and ADAS results, non-conformance and takt history for retraining and benchmarks.
Spec, IATF, functional-safety, cyber and human-approval policy enforced before any write; supportable on-prem or plant-edge deployment for regulated OEM plants.
Chassira writes to production equipment. Every capability is scoped, every write is policy-checked against the qualified envelope, safety-critical parameters are locked, and every action is written to an append-only audit log the plant owns.
| Item | Scope | Status |
|---|---|---|
| SOC 2 Type I | Cloud control plane | QUEUED Planned, first 6 months, with an IATF and functional-safety-aware review |
| SOC 2 Type II | Cloud control plane | QUEUED Planned, 6 to 12 months |
| SSO and RBAC | Console and approvals | QUEUED Planned, 6 to 12 months |
| IATF 16949 and APQP assurance packages | Disposition and traceability evidence | QUEUED Planned, 6 to 12 months; not a certification claim |
| ISO 26262-aware control | Functional-safety requirements cited in reasoning and gates | RUNNING In the design; not a certification claim |
| Standards-cited reasoning | IATF 16949, APQP and PPAP, torque specs, recall procedures | RUNNING In the design, every answer cites its source |
| Per-tenant and program isolation | Designs, control plans, models, memory | RUNNING In the design from day one |
| On-prem and plant-edge deployment | Regulated OEM plants | RUNNING Design option |
The questions plant managers, launch engineers and quality leads ask in the first meeting.
Yes, but only within an explicit tag allow-list with per-tag rate and magnitude limits, inside the qualified envelope, and only at the autonomy level your plant has set. Design partners start in shadow mode, where Chassira writes nothing. Weld-schedule changes wait for a named body-shop engineer, and torque specifications, battery-joint parameters and end-of-line test limits and verdicts are read-only to every agent, always.
Control returns to the press, robot, booth or MES controller's last known-good state. Chassira is a supervisory layer on top of the safety-rated PLCs and interlocks you already run, never a replacement for them, so an outage degrades the line to its current way of running, not to a stop.
Advisory recommendations start from the plant twin and the control plan in the first weeks. Weld, geometry, paint, torque and end-of-line prediction improve as line history and labelled outcomes accumulate per program. The compute plan sizes onboarding at 150 to 600 GPU-hours per plant and 50,000 to 250,000 synthetic variants per program, all design targets.
Only if you choose cloud training. Designs, control plans, specs, process models and quality records are isolated per tenant and never used to train another OEM's models. On-prem and plant-edge deployment are available for regulated plants, and federated cross-plant learning is planned and waits on data-rights agreements.
You are, the same as with any control strategy, which is why every write is policy-checked, bounded, logged and reversible, and why anything crash-safety or IATF relevant waits for a named approver. The audit log records the request, the evidence, the envelope applied and the human decision.
A paid pilot on one process or area at the Process tier, scoped to a single wedge with a success metric written into the order form: weld or geometry rework, downtime minutes, paint defect or repaint rate, warranty or PPM, or launch quality. Shadow mode first, then assist mode with your engineer approving each move, then bounded autonomy on low-risk loops once ROI and safety gates are met.
The SDK and CLI are design previews. Tell us what you would build and we will tell you where it stands.
↑↓ navigate↵ openesc close