How to build an agent factory, stage by stage
An agent factory is the old development lifecycle closed into a loop. Each stage, why it exists, what breaks without it, and which one to build first.
- agent factories
- control loops
- SDLC
Blog
Notes from a research programme on the autonomous enterprise: accountability, decision records, boundaries, the off switch and why a machine cannot own anything.
Who owns what an agent factory learns
The valuable part of an agent factory is not the model but what it learns about your firm. Before choosing a vendor, check what stays yours when you leave.
- agent factories
- sovereignty
- vendor lock-in
Agent factories will not stay in software
Software teams now run work as a loop of agents with people at the decision points. Finance, purchasing and marketing come next, on one condition.
- agent factories
- closed loops
- accountability
The supervisor has to sit outside the system it watches
NVIDIA's new agent security design puts the monitor on a separate processor, outside the host. It is the industrial form of an off switch the system cannot reach.
- supervision
- agents
- isolation
Auditing a system that changes faster than its auditor
An audit assumes a stable object. A self-improving system has changed by the time the report is written, so the audit has to move from behaviour to invariants.
- audit
- attestation
- invariants
An autonomous firm's budget: money the loop cannot spend
A loop that can buy, rent compute and pay suppliers needs a spending bound it cannot edit. The budget is the first constraint an owner can actually read.
- budget
- spending limits
- economic behaviour
A decision log nobody can replay is not a ledger
Most systems keep logs for an engineer on call tonight. An autonomous one needs a record that a stranger can replay a year later, and that is a different design.
- audit
- decision records
- traceability
A system that improves itself needs a yardstick it cannot edit
The third loop of an autonomous operation changes its own configuration to do better. Who defines better, and where that definition lives, decides if it works.
- objectives
- metrics
- Goodhart's law
Autonomy is defined by what the loop cannot touch
Giving a system more autonomy is mostly a decision about what it is forbidden to do and who enforces that. A boundary the loop can edit is only a suggestion.
- permissions
- limits
- blast radius
Software can be rolled back. A forklift cannot
Extending an autonomous loop from records to machines turns wrong data into moving mass, and the safety case has to be written before the loop is connected.
- robotics
- physical systems
- safety case
What a sandbox market can and cannot tell you
Studying economic behaviour without real counterparties is the only defensible option today. It pays to be exact about what that leaves untested.
- simulation
- sandbox
- pricing
The last human in the loop is a signature
Most automation programmes stop at the same place, and it is not a technical boundary. Somebody has to be accountable, and accountability needs a person.
- automation
- accountability
- approval
Self-healing is not a feature, it is a different architecture
Bolting recovery onto a system that was designed to be repaired by people produces automation that fails in more interesting ways, not fewer.
- self-healing
- reliability
- control loops
Can a machine own anything?
Legal personhood for autonomous systems gets discussed as though it were arriving. It is not, and the reasons are more interesting than the proposals.
- legal personhood
- liability
- regulation
An autonomous system needs an off switch it cannot reach
The interesting property of a stop is not that it exists. It is that the thing being stopped has no path to observe, delay or depend on it.
- kill switch
- safety
- supervision
Why we call it an organism, and what that does not mean
Biological language is useful for describing closed loops and self-repair, and it is the fastest route to claims nobody can support. Both are worth saying.
- terminology
- cybernetics
- claims