O-AGENT
Decide and actThe only part that acts, and the only part that has to earn the right to.
O-AGENT reads across imagery, telemetry and the record, decides what deserves attention, and does the work that follows. Every action runs inside a limit you set, escalates when it is not confident, and leaves a trace you can audit.
O-AGENT reads the whole model — imagery, telemetry and record — decides what should happen next, and acts only inside the limits you set, leaving a trace behind every action.
Nobody's problem is a shortage of dashboards.
A portfolio team can already see everything. What they cannot do is read four systems for two hundred sites every week, decide which of eleven thousand findings matter, and get the right eight of them in front of a crew before Friday. That is not an analytics problem. It is a decisions-per-week problem, and adding another chart does not move it.
Reading, weighing, and knowing when to stop
What actually happens between eleven thousand findings and the handful of jobs a crew is sent to do on Monday.
- 01
It reads all three at once
Imagery, telemetry and the maintenance record, on the same asset, in the same pass. A person can do this for one string. Nobody does it for two hundred sites every week.
o-agentIt reads all three at once1/4 - 02
It ranks by what the evidence supports
A hotspot corroborated by six weeks of current drift on the same string is not the same finding as a hotspot on its own, even though both look identical in a report.
o-agentIt ranks by what the evidence supports2/4 - 03
Every call carries a confidence
Not a label, a number, and one you set the bar for. The thermal signature without a telemetry signature to back it sits at 61%, and it says so.
o-agentEvery call carries a confidence3/4 - 04
Below your bar, it stops and asks
Two dispatch on their own. The one it is unsure about goes to a person, with the reasoning attached. An agent that never declines is one nobody will let near a substation.
o-agentBelow your bar, it stops and asks4/4
From reading to acting
Six things an agent does, and the last one is the reason you can allow the other five.
Reads the whole model
An agent that can only see the inspection feed is guessing. These read imagery, telemetry and work history against the same assets, which is the difference between a detection and a decision.
Triage by real severity
Ranked by what it means for the asset and the portfolio, not by detection confidence. A high-confidence hotspot on a string that is already isolated is not urgent.
Plans the next capture
Works out what the record is missing or has gone stale on, and builds the flight plan to fill it, rather than flying the same grid every quarter regardless.
Predicts what fails next
Uses telemetry drift and work history together to say what is likely to fail, when, and what it would cost to wait.
Raises and dispatches work
Creates the work order in your system with the evidence attached, assigns it, and closes the loop when the job is done.
Leaves a trace
Every decision records what it read, what it concluded, what it did, and which threshold let it. Exportable, and readable by someone who was not there.
The agents
Not one general assistant. Each of these does a specific job, reads specific things, and has its own limit on what it may do without a human.
Inspection planner
Reads
Asset register, last pass dates, weather windows
Does
Builds and schedules the next capture campaign
Runs unattended up to
Scheduling. Flight authorisation stays with your pilot.
Anomaly triage
Reads
Classified findings, telemetry, work history
Does
Ranks findings by real severity and merges duplicates
Runs unattended up to
Ranking and merging. Critical findings always surface to a person.
Maintenance scheduler
Reads
Findings, drift signals, parts, crew availability
Does
Raises and schedules work orders against available resource
Runs unattended up to
A cost ceiling you set. Above it, it proposes rather than raises.
Compliance reporter
Reads
Inspection evidence, certificates, audit trail
Does
Assembles the periodic report in the format the regulator wants
Runs unattended up to
Drafting. Submission is always a human action.
Field copilot
Reads
The asset's full history, on a phone, offline
Does
Answers a technician's question at the asset and captures what they find
Runs unattended up to
Read and capture. It does not close its own jobs.
Where your team already works, and where your systems do
Nobody adopts a new console to ask one question. O-AGENT answers on the surfaces your operations team is already in, and exposes the same agent to your own software over MCP, a REST API, webhooks and SDKs. Same reasoning, same limits, same decision trace, whoever is asking.
Chat
- SSlack
- MTMicrosoft Teams
- Telegram
- Discord
- SignalPlanned
- Gmail
- OOutlook
- SMShared mailbox
Voice and SMS
- SSMS
- WCWeb chat
- VVoicePlanned
Programmatic
- MCP server
- REST API
- WWebhooks
- Python SDK
- TypeScript SDK
- GraphQLPlanned
A Tuesday evening, on the channel it came in on
What is going on with Binh Thuan 02? Output is down about 4% today and I am on shift until ten.
Three strings on inverter 04 are underperforming. B-07-14 has a bypass diode failure from Tuesday, confirmed by six weeks of current drift. The other two are soiling, and rain is forecast Thursday.
Raise the diode one. Leave the soiling.
WO-2214 raised against string B-07-14, Thursday 06:00 window, junction box kit confirmed in stock. Field team 2 notified on WhatsApp. I did not touch the soiling findings.
Governed autonomy
The reason an operator will let this near a substation is not that the model is good. It is that the limits are explicit, the escalation is automatic, and the trace exists whether or not anything went wrong.
Per-action limits
Autonomy is set per action, not per product. Raising a work order and dispatching a crew are different permissions, and you set each.
A confidence bar
Below the bar you set, an agent stops and asks. It does not act at low confidence and explain afterwards.
Automatic escalation
Anything critical, anything novel and anything above a cost ceiling goes to a named person, with the reasoning attached.
Full decision trace
What it saw, what it weighed, what it decided, what it did. Exportable for audit and readable without a data scientist.
Your region, your tenancy
Data stays where you require it to stay, and the model does not train on it across customers.
Reversible by default
Actions an agent takes can be undone by a person, and the undo is recorded the same way the action was.
- 01O-EYEDetectInspection & digital twin
- 02O-ARCAnalyseSensors, IoT & telemetry
- 03O-OPSOperationsThe system of record
- 04actsO-AGENTDecide & actThe agentic layer
Every action writes back to the record, so the model sharpens with every job.
Questions
What happens when it gets something wrong?
The trace shows what it read and which threshold allowed the action, so you can see whether the reasoning or the limit was the problem. Actions are reversible, and the limit can be tightened immediately.
Can we run it read-only at first?
That is the normal way to start. Agents propose, a person accepts, and you watch the agreement rate for a few cycles before letting anything run unattended.
Does it train on our data?
Not across customers. Your data stays in your tenancy and your region, and improvements from your usage stay yours.
Who is accountable for an agent's decision?
You are, which is exactly why the limits are yours to set and the trace is written for a human reader. We do not present this as autonomy that transfers responsibility.