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.
Agent
Developers
Typed tools. Declared limits. Replayable runs.
The Draweon SDK is a design preview of how an agent is defined, bounded and run: tools with unit-aware ranges and rate limits, autonomy policy as code, deterministic replay of any historical run against a new model, and a local twin harness. The API will change before release.
Illustrative scenario: EX-3 · 2.5 mm² class 2 PVC building wire, changeover, no spark faults
Concepts
Five objects
Everything in the SDK is one of these.
A connector call with a schema
Tools name tags, units and whether they are read-only. Write tools carry Limits.
Toolread_only
Rate and magnitude
max_step per period, absolute floors and ceilings, enforced at call time by the policy engine and again at the OT boundary.
Limit
L1 to L4
The level the plant has set for this agent, this tag, this shift. Approvals are routed to named humans above the threshold.
Autonomy.L1Autonomy.L3
A goal becoming a graph of steps
Start, stream, replay. Every step has an agent, a tool, a duration and a status, and the whole record is kept per reel.
run.stream()run.replay()
verify=twin
Simulate a move on the as-run line twin before it is written. Required for any agent configured with it.
twin.simulate
Developers
Define an agent, bound it, run it
The Draweon SDK is typed Python. Tools are declared with schemas and limits and the policy engine enforces them at call time, not in a review meeting. This is a design preview; the API will change before release.
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
# Design preview. Bound the extrusion agent to line EX-3.
from draweon import Agent, Tool, Limit, Autonomy
screw = Tool(
name="plc.write_setpoint",
tags=["EX3.SCREW_RPM", "EX3.ZONE3_T"],
limits=[Limit(max_step="2%", per="30s")],
)
extrude = Agent(
id="agent.insulate_jacket",
goal="wall 0.80 mm nominal, min 0.62, concentricity >= 0.90",
tools=[screw, Tool("gauge.read_wall", read_only=True)],
# bounded writes; humans still gate ramps
autonomy=Autonomy.L3,
# simulate on the twin before every write
verify="twin",
)
run = extrude.start(line="EX-3", construction="BW-2.5-C2-PVC")
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. Design preview.
$ draweon run "changeover EX-3 to BW-2.5-C2-PVC" --autonomy L3
→ plan composed 10 steps · 1 approval gate
→ twin.simulate 36 candidates · best #22 · spark risk 0.04
→ policy.evaluate 8 writes permitted · 1 held for human
→ executing draw.tension ..... ok 1m48s
→ executing anneal.current ... ok 54s
→ executing strand.lay ....... ok 1m12s
→ executing extrude.wall ..... ok 2m36s
! approval required line_speed 1100 → 1250 m/min
→ approved shift process engineer · 06:41:52
→ scenario complete spark faults 0 · reel R-30417 released
$ draweon runs show run_3d7a1e --format genealogy
Run timeline · scenario
One scenario run, step by step
Every Draweon run is an ordered, inspectable sequence. This is run_3d7a1e on EX-3 · 2.5 mm² class 2 PVC building wire. Scenario completes in 10 steps with one approval gate. Wall settles at 0.80 mm nominal, minimum 0.66 mm, concentricity 0.93, no spark faults, reel R-30417 released with full genealogy.
run_3d7a1e
EX-3 · 2.5 mm² class 2 PVC building wire
illustrative scenario
SUCCEEDED
-
01
ingest.order
SUCCEEDED
0.6 s
Plant Orchestrator · Pulled the construction spec for 2.5 mm² class 2 PVC building wire from MES: 7 × 0.67 mm conductor, 0.80 mm nominal wall, 0.62 mm minimum spot wall, 6 kV spark test, 48 km order. Locked the target envelope.
-
02
twin.simulate
SUCCEEDED
41 s
Line Twin · Simulated 36 candidate recipes on the as-run line twin for draw schedule, lay length and extrusion setpoints, ranked on wall variance, concentricity, predicted spark risk and polymer give-away.
-
03
draw.tension
SUCCEEDED
1 m 48 s
Draw-and-Anneal · Capstan tension on RB1 stepped down 6% as the die-wear estimate crossed its threshold. Diameter held at 0.670 mm ± 0.004 mm and the break-risk score fell from 0.31 to 0.09.
-
04
anneal.current
SUCCEEDED
54 s
Draw-and-Anneal · Annealer current trimmed against the live conductivity reading to hold the softness target without over-anneal through the speed change.
-
05
strand.lay
SUCCEEDED
1 m 12 s
Strand-and-Cable · Lay length set to 28 mm on bunching line ST2 with payoff tension balanced across 7 spools. Roundness index held above 0.97 into the extrusion payoff.
-
06
extrude.wall
SUCCEEDED
2 m 36 s
Insulate-and-Jacket · Screw speed and zone 3 temperature re-cut against melt pressure and material-lot moisture. Wall settled at 0.80 mm nominal, minimum 0.66 mm, concentricity 0.93.
-
07
inspect.fuse
SUCCEEDED
continuous
Defect-and-Inspect · 6 cameras, the laser gauge, the eddy-current head and the spark tester fused at the edge. One neckdown precursor at 12,480 m flagged and cleared by a screw-speed trim before it reached the reel.
-
08
approve.human
APPROVAL
38 s
Plant Orchestrator · Line-speed ramp 1,100 to 1,250 m/min exceeded the site autonomy threshold. Held for the shift process engineer and approved at 06:41:52.
-
09
reel.change
SUCCEEDED
2 m 05 s
Robot-and-Handling · Reel change on take-up TU3 planned and executed by the handling robot. Payoff aligned and tension re-established inside the changeover window.
-
10
reel.release
SUCCEEDED
1 m 10 s
Quality-and-Conformance · Reel R-30417 released: genealogy written (conductor lot, draw and anneal record, lay, wall and concentricity trace, defect map, approvals) and checked against the construction spec.
Tool calls · scenario
The agent shows its work
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.
- 06:30:12mes.read_order({ line: "EX-3", construction: "BW-2.5-C2-PVC" })
- 06:30:12ok spec locked · wall 0.80 mm nominal · min 0.62 mm · spark 6 kV
- 06:30:13twin.simulate({ candidates: 36, horizon: "14 min" })
- 06:30:54ok best recipe #22 · wall var 0.018 · spark risk 0.04 · give-away 2.1%
- 06:30:55policy.evaluate({ recipe: 22, autonomy: "L3" })
- 06:30:55ok 8 writes permitted · 1 write requires human approval
- 06:30:56plc.write_setpoint({ tag: "RB1.CAPSTAN_T", delta: "-6%" })
- 06:32:44ok diameter 0.670 mm ±0.004 · break risk 0.31 → 0.09
- 06:32:45plc.write_setpoint({ tag: "ANN1.CURRENT", mode: "hold_conductivity" })
- 06:33:39ok conductivity on target · softness in window
- 06:33:40plc.write_setpoint({ tag: "ST2.LAY", to: "28 mm", spools: 7 })
- 06:34:52ok roundness 0.97 · payoff tension balanced
- 06:34:53plc.write_setpoint({ tag: "EX3.SCREW_RPM", zone3: "+4 °C" })
- 06:37:29ok wall 0.80 nominal · min 0.66 · concentricity 0.93
- 06:37:30sensors.fuse({ cams: 6, laser: true, eddy: true, spark: true })
- 06:39:04precursor: neckdown risk 0.58 · 12,480 m · zone 3
- 06:39:05plc.write_setpoint({ tag: "EX3.SCREW_RPM", delta: "+1.5%" })
- 06:39:21ok neckdown risk 0.58 → 0.06 · wall recovered
- 06:39:22approval.request({ action: "line_speed", to: "1250 m/min" })
- 06:41:52ok approved by shift process engineer · logged
- 06:41:53isaac.plan_path({ task: "reel_change", takeup: "TU3" })
- 06:43:58ok reel changed · payoff aligned · tension re-established
- 06:43:59mes.log_reel({ reel: "R-30417", genealogy: "full" })
- 06:45:09ok scenario complete · 10 steps · spark faults 0 · released
Reasoning trace
Plan, thought, action, observation, exactly as the orchestrator would record it.
GoalChange line EX-3 over to 2.5 mm² class 2 PVC building wire, ho…
Change line EX-3 over to 2.5 mm² class 2 PVC building wire, hold 0.80 mm nominal wall and concentricity, and finish the run with zero spark faults.
PlanSimulate the changeover on the twin first. Lead with drawing a…
Simulate the changeover on the twin first. Lead with drawing and annealing so the conductor is right before stranding, then set lay, then re-cut extrusion against the live wall reading. Fuse inspection continuously. Hold the line-speed ramp for a human.
ThoughtThe standing recipe raises line speed and screw speed together…
The standing recipe raises line speed and screw speed together. On the twin that produces a wall dip at zone 3 because the melt pressure lags the speed step. Sequencing the zone 3 temperature move ahead of the speed ramp removes it.
Actiontwin.simulate(candidates=36) returns recipe #22: zone 3 up 4 °…
twin.simulate(candidates=36) returns recipe #22: zone 3 up 4 °C before the ramp, screw speed trimmed to the live wall reading, lay 28 mm.
ObservationRecipe #22 predicts wall variance 0.018 mm, spark risk 0.04 an…
Recipe #22 predicts wall variance 0.018 mm, spark risk 0.04 and give-away 2.1%. Three candidates scored lower on give-away but pushed predicted minimum wall under the spec floor and were discarded.
ActionExecute recipe #22 under autonomy level L3: eight setpoint wri…
Execute recipe #22 under autonomy level L3: eight setpoint writes permitted, the line-speed ramp routed to the shift process engineer.
ObservationNeckdown precursor at 12,480 m at 06:39:04. A 1.5% screw-speed…
Neckdown precursor at 12,480 m at 06:39:04. A 1.5% screw-speed trim cleared it in 16 s with no reel quarantined.
OutcomeScenario complete. Wall 0.80 mm nominal, minimum 0.66 mm, conc…
Scenario complete. Wall 0.80 mm nominal, minimum 0.66 mm, concentricity 0.93, zero spark faults. Reel R-30417 released with full genealogy.
Autonomy
Four levels, set per agent and per tag
A plant does not go from manual to unattended in one step. Draweon makes the level explicit, auditable and reversible at any time. The first release runs shadow mode, then assist mode, then graduated autonomy.
| Level | What the agent does | What the human does | Where it sits in a pilot |
|---|---|---|---|
| L1 · Advisory | Recommends setpoints and explains why, with citations | Enters every change manually | Shadow mode, weeks 1 to 3 |
| L2 · Supervised | Proposes a write; it executes on approval | Approves each write in the review console | Assist mode, weeks 4 to 8 |
| L3 · Bounded | Writes inside tag, rate and magnitude limits | Approves ramps, die changes and reel release | Graduated autonomy, from week 9 if the plant is satisfied |
| L4 · Unattended | Runs the envelope without prompting | Sets the envelope; reviews the shift record | Planned; after qualification on the construction |
Accelerated computing
Physical AI, at the die train
Conductor moves at over a thousand metres a minute. Perception has to be local, deterministic and fast, so Draweon runs GPU inference at the plant edge and keeps training, simulation and optimisation in the cloud or on-prem.
Edge perception
4 to 12 cameras per line plus laser, eddy-current and spark streams fused at the line. Sub-50 ms alerting and sub-250 ms bounded control are the design targets. Jetson Thor is planned for higher sensor density.
Model serving
Vision, time-series, anomaly and quality models behind one orchestrator at the plant edge, with canary rollout by line and fallback to the previous validated version.
Training
Defect classifiers, sensor-fusion models, time-series prognostics and cable-process reasoning models trained on reel telemetry and engineer corrections. Dedicated training capacity is planned.
Line twin
As-run drawing, stranding, extrusion, thermal and tension simulation. Physics-informed surrogates are planned so recipes can be optimised at production cadence.
Synthetic data
Rare spark, pinhole, eccentricity, neckdown and break variants generated and validated against real gauge, spark and vision distributions before promotion. Planned.
Optimisation
Line sequencing, reel changeovers, energy allocation, takt and constrained setpoint proposals.
Robot handling
Spool, reel and payoff and take-up motion planned and validated in simulation before it runs next to moving conductor.
Guardrails
An agent that can move a screw speed needs a leash
Draweon 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 writes only to an explicit tag allow-list, inside per-tag rate and magnitude limits.
- Policy engine before every write. Autonomy level, shift, construction, interlock state and operator presence are evaluated before a setpoint moves.
- Human-in-the-loop gates. Anything above the site threshold, such as line-speed ramps, die changes and reel release, waits for a named approver.
- Immutable audit log. Append-only and exportable. Every action records source signals, retrieved procedures, predicted impact, allowed range, rollback condition, approver and reel scope.
- Hard fallback. Loss of the edge node, the network or a model returns control to the line controller's last known-good state.
- Tenant and IP isolation. Construction recipes, defect libraries and process memory never cross a customer boundary. On-prem deployment is available for sensitive producers.
Compliance posture
| Item | Scope | Status |
|---|---|---|
| SOC 2 Type I | Cloud control plane | QUEUED Planned, first 6 months |
| 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 |
| Standards-cited reasoning | UL, IEC, IPC, USCAR documents | RUNNING In the design, every answer cites its source |
| Per-tenant isolation | Recipes, models, memory | RUNNING In the design from day one |
| On-prem and plant-edge deployment | Sensitive producers | RUNNING Design option |
FAQ
Straight answers
The questions plant directors and process engineers 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 plant has set. Level 1 is advisory: Draweon recommends and a human enters everything. Design partners start there, in shadow mode, before supervised writes.
Control returns to the line controller's last known-good state. Draweon is a supervisory layer on top of the PLCs and gauges 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 physics-based twin and the construction spec in the first weeks. Defect prediction and break prediction improve as line history and labelled events accumulate per construction. The first training sets are sized at 5 to 20 million labelled frames and events, a design target.
Only if you choose cloud training. Recipes, process models and defect libraries are isolated per tenant and never used to train another customer's models. A fully on-prem deployment at the plant edge is available for 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 paid pilot on one line, scoped to a single workflow with a baseline agreed in writing. Weeks 1 to 3 are connection and shadow-mode observation, weeks 4 to 8 are assist mode with engineer approval, and bounded autonomy follows if the plant is satisfied with the recommendations.
Get started
Ask for early access
The SDK and CLI are design previews. Tell us what you would build and we will tell you where it stands.