One identity for people and agents: who may do what
Agents acting under a shared service account make every action anonymous. Why an agent needs its own identity, delegated from a person, with scopes that expire.
- identity
- permissions
- agents
Blog
Notes from building an agentic operating system: ontology, agents, data, sovereignty and models, and the places where the automation deliberately stops.
The live state and the record: two memories of a company operating system
A company operating system needs two memories: the live state modules change every second, and the record of every event, proposal and decision for the audit.
- record
- lakehouse
- audit
An agent that writes to the ERP: the line between read and write
Reading from a system of record is cheap and safe. Writing back is where integrations fail. What has to exist before an agent may write, and where Hurukan draws its line.
- system of record
- ERP
- agents
Apache Paimon: one state table for the operator, the planner and the agent
How Apache Paimon keeps a table that a stream updates thousands of times an hour, how its merge engines express ontology rules, and what it costs to run.
- Apache Paimon
- LSM tree
- merge engine
One record, readable by the tools you already have: XTable and UniForm
Apache XTable and Delta UniForm let one record be read as Delta, Iceberg or Hudi without a copy. What they translate, what they do not, where lock-in sits.
- Apache XTable
- Delta UniForm
- table formats
Android XR as an enterprise endpoint for on-device agents
Android XR brings Gemini, a camera and agents onto glasses. Before a company rolls them out: what runs on the device, what at the vendor, what in the company's own system.
- Android XR
- on-device agents
- smart glasses
XREAL Aura on the shop floor: the system sees the workstation
XREAL Aura puts a camera, a display and Android XR in front of a frontline worker. What it takes for what the glasses see to end up in the company's own record.
- XREAL Aura
- Android XR
- smart glasses
Apache Hudi: the coexistence phase, change data from the old ERP into the record
Modules take over one process at a time; until then the old ERP keeps writing. How change data capture and Apache Hudi keep the record current meanwhile.
- Apache Hudi
- change data capture
- coexistence
The state domain agents read: Paimon for the work queue, stock and approvals
A domain agent needs the company's state as of now, not a nightly copy. How the Data layer keeps it in Apache Paimon, and why agents only propose.
- Apache Paimon
- domain agents
- live state
Delta Lake: when the company stands on Microsoft
OneLake stores tables as Delta Parquet. What the Delta log gives a company's record, what Microsoft Fabric manages for you, and what remains yours to decide.
- Delta Lake
- Microsoft Fabric
- OneLake
Apache Iceberg: the record you can take with you
Iceberg keeps a table as a tree of immutable metadata with one pointer in a catalogue. Why so many vendors read it, and why the catalogue decides sovereignty.
- Apache Iceberg
- table format
- catalogue
The audit trail is a table: Delta Lake, Iceberg or Hudi
Who changed what, and what did the agent see when it proposed it? Delta Lake, Iceberg and Hudi answer differently. Choose by catalogue and infrastructure.
- table formats
- Delta Lake
- Apache Iceberg
Small models on your own hardware: when they are enough
Not every task needs a frontier model in someone else's data centre. Small open-weight models on your own hardware handle many narrow jobs well. How to tell which ones.
- small language models
- open-weight models
- on-premises
Apache Parquet: how the record is laid out so superintelligence can read the whole company
Superintelligence reads the record constantly, on the company's own hardware. Parquet's layout sets what each question costs: row groups, statistics, codecs.
- Apache Parquet
- the record
- columnar storage
Parquet and Avro: the event between modules and the record superintelligence reads
An event travels between modules as a row with a contract and rests in the record as columns. Why Avro carries the first and Parquet the second.
- Apache Parquet
- Apache Avro
- event contracts
Retrieval over the ontology beats retrieval over documents
Retrieval over a pile of documents finds text that looks relevant. Retrieval over a model of the company finds the facts that are true now, together with their relations.
- RAG
- retrieval
- ontology
A system of record cannot act
Business software was designed to hold an accurate account of what already happened. That design decision is why a company's speed is set by how fast its staff can type.
- system of action
- ERP
- automation
Apache Avro: the contract between modules
Every event between modules is a contract. Apache Avro writes it into the data, and a schema change becomes a proposal a person approves and the system records.
- Apache Avro
- schema registry
- event contracts
One definition of revenue for every report and every agent
When finance, sales and the board each compute revenue their own way, agents inherit the confusion at machine speed. A semantic layer fixes each definition once, upstream.
- semantic layer
- metrics
- data governance
The integration tax
Every company knows what its software costs. Almost none can say what the space between the systems costs, and it is usually the larger number.
- integration
- data silos
- cost
Agents, RPA or workflows: which automation for which job
Workflow engines, RPA bots and language-model agents solve different problems. Choosing by fashion gives brittle bots and unpredictable agents. How to choose by the job.
- RPA
- workflow
- AI agents
Intelligence has a lawful meaning and an unlawful one
Borrowing intelligence vocabulary for a business product is easy and mostly harmless. Borrowing the practices those words describe is a criminal offence.
- OSINT
- GDPR
- boundaries
Voice for hands-busy work: conversation on the shop floor
When both hands hold a scanner, a part or a tool, a screen is the wrong interface. What voice has to do to be useful in a warehouse, a workshop or on a service visit.
- voice interface
- speech recognition
- warehouse
Code written by agents still gets reviewed
An autonomous software factory is a real advantage and the easiest thing in this field to oversell. Here is exactly where the automation stops.
- AI engineering
- CI/CD
- review
When the company changes, the ontology has to change with it
A company adds a service line, splits a region or buys a competitor, and its model of the business goes stale. How to change an ontology without breaking what depends on it.
- ontology
- versioning
- change management
Compute is a market, not a subscription
AI work is bursty and hardware sits idle. A grid where machines trade capacity is an old idea whose obstacles were never technical, and there is no token in it.
- distributed compute
- grid
- settlement
Which data must never leave the building, and how to decide
Treating every record as secret makes the rule unworkable, and treating none as secret makes it pointless. A practical way to classify data before any model touches it.
- data classification
- GDPR
- trade secrets
Classical ML still beats an LLM at forecasting demand
Asking a chat model how much stock to order produces a fluent number, not a forecast. Demand forecasting is still a job for classical models, a baseline and a backtest.
- demand forecasting
- machine learning
- backtesting
Where large language models earn their keep in a company
Language models are strong at reading, drafting and translating between people and systems, and weak at arithmetic, recall and final decisions. A map of where to use them.
- LLM
- use cases
- document processing
The dashboard nobody opens: BI that answers questions
Dashboards are built for the questions someone had last year. Managers have new ones every week. Analytics should start from the decision and the question, not the chart.
- business intelligence
- dashboards
- metrics
Hyperautomation fails at the handover, not the task
Each automated step works in the demo, and the process still stalls, because nobody designed what happens between the steps: the handover, the exception and the owner.
- hyperautomation
- process design
- exceptions
A data platform is not a data lake with a dashboard
Copying every system into one store and pointing a BI tool at it builds a bigger pile, not a platform. A platform is defined by meaning, ownership and a way back to action.
- data platform
- data lake
- data ownership
Your company already has an ontology. It lives in people's heads
Every company runs on a shared model of customers, orders and exceptions that nobody wrote down. How to find where that model lives and how to get it out of people's heads.
- ontology
- tacit knowledge
- business rules
A chat window is not a conversational interface
Adding a chat box to business software gives people a new place to type, not a new way to work. A conversational interface needs context, permissions and the right to act.
- conversational interface
- chat assistant
- agents
Sovereignty is an architecture decision, not a hosting choice
Moving servers to a local data centre does not make a company sovereign. Who holds the keys, the model of the business and the exit does, and that is decided in the design.
- data sovereignty
- vendor lock-in
- encryption keys