Architecture
The system behind every operator.
An operational system that runs on the systems you already use: a synchronized data layer, an operator that orchestrates over 300 subagents, and an action model trained per company.
Running an operational role end to end takes more than a model. It takes infrastructure: data that stays current, orchestration that holds hundreds of moving parts, and a model fit to how the operation works.
How an operator runs.
An operator works across the layers you already own — from your systems to a final action — orchestrating over 300 subagents in between.
Your systems
The tools you already use — TMS, ERP, WMS, email, customer portals. Operators read and act through a synchronized data layer, with no migration.
The operator
One controller orchestrates over 300 specialised subagents, each narrow and accountable, to run an entire area of work end to end.
Your model
Each company runs its own action model — a small language model tuned on its operation, and never shared across customers.
What the system is built on.
- 01 · Data layer
A synchronized layer over your systems.
Each deployment runs on a dedicated data layer kept in sync with the systems you already use — TMS, ERP, WMS, email, customer portals. Operators read and act through this layer rather than the source systems, working from a current view with no migration.
- Per customer
- Change-data-capture sync
- 150+ connected systems
- Agent-ready views
Data layer
4 systems- TMSSynced
- ERPSynced
- WMSSynced
- EmailSynced
One synchronized layer - 02 · Orchestration
One operator orchestrates over 300 subagents.
An operator is an orchestration system, not a single model. A controller routes each unit of work across more than 300 specialised subagents — some that reason, some that act — each one narrow and accountable. The decomposition lets one operator run an entire area of work end to end.
- Hub-and-spoke
- 300+ subagents
- Reasoning + automation
- Per-action accountability
Operator312 subagents- Validate HS code#a12
- Route exception to carrier#a47
- Reconcile invoice #4188#a63
- Update delivery ETA#a71
- 03 · Expertise capture
Operational expertise is captured from your team.
The constraint in operational AI is tacit knowledge — the judgment calls, exceptions, and rules that are never written down. It is captured through structured elicitation drawn from cognitive and behavioural science, then encoded into the operator.
- Structured elicitation
- Tacit-knowledge capture
- Cognitive + behavioural methods
Expertise capture
3 rulesElicitation
“If a load’s over five grand, I always get a manager to sign off…”
- Hold loads over $5k for sign-off
- Reroute on >2h port delay
- Escalate customs mismatch
- 04 · Action models
Each company runs its own model.
Operators do not share a single general model. Each company runs its own action model — a small language model tuned on its operation — so decisions reflect how that business works. The model improves as the operator runs and is not shared across customers.
- Small language model
- One per company
- Improves with use
- Self-hostable
Action model
tuning96.4%
Agreement with team
step
2,480
One model · this company
- 05 · Graduated autonomy
Autonomy is granted in stages.
An operator starts in shadow mode — proposing actions rather than taking them — inside guardrails that define what it can touch, what needs sign-off, and when to escalate. As its decisions match your team’s, it is granted more autonomy. Every action is logged and reversible.
- Shadow to autonomous
- Guardrails
- Approval thresholds
- Full audit trail
Autonomy
level 3 / 5ShadowSupervisedAutonomousApprovals
Booking #2245autoRate $2,400 — CHI→DALreview
How an operator goes live.
- 01
Discovery
We capture how the work is done today — the decisions, exceptions, and judgment that are not written down.
- 02
Build
The operator is assembled against your live systems: data layer connected, subagents configured, guardrails set.
- 03
Deploy
It runs in shadow mode, proposing actions, so the work can be checked before anything is executed.
- 04
Learn
Each approval or correction updates the model. As it proves out, the operator is granted more autonomy.
The shape of the system.
300+
Subagents orchestrated by one operator
1
Action model per company
150+
Systems connected through one data layer
100%
Of operator actions logged and reversible