Problems you already have
Someone in your organisation shipped an agent this month. It has access to something, it ran last night, and if you asked today who approved that, you would be waiting for an answer. Each film takes one of those questions and shows it answered on the running product, in under two minutes.
If you watch one, watch this
The question every other film here is a version of.
Start here
Nobody can tell you what your agents are allowed to do
Agents are being built across your organisation, probably by people who do not know about each other. The question is not what they were built to do. It is what they are allowed to do today, and who decided.
Building and running agents
For the people who have to ship them, and the people who sign them off.
Delivery and architecture
The life of an agent, end to end
Seven stages from proposed to retired, each with a gate. Change the model, the tools or the rules and the gate that checked them reopens.
Delivery leads
Governance that does not stop delivery
When governance means waiting, teams stop telling you what they are building. A gate that names the one requirement that is unmet, rather than simply refusing.
Data owners
Who decided this agent could read customer data?
One decision, taken by a named seat, with the reason kept on the record — and the agent’s class and every permission after it computed rather than typed.
Proving it, and reporting it
For everybody who is eventually asked to show their working.
Risk, audit and compliance
The day somebody asks you to prove it
A screen full of green is not evidence. Every figure names its source, or says it has none — and a zero it cannot stand behind is never printed as a zero.
Programme and portfolio leads
Adoption and metrics you can put your name to
Adoption, quality and cost, each paired with the measure that keeps it honest, and every figure labelled measured, sampled, assumed or not computable.
Whoever reports upwards
What you put in front of the board on Monday
Not a demonstration — a page they can take away and be held to. Built for a scope and a period, with a control that fired counted as evidence the control works.
Internal audit
The evidence an auditor can check without you
A seat that reads everything and changes nothing, a record held so that alteration is detectable, and a plain account of what reconstruction does and does not mean.
Setting it up so it holds
Policy, tools, keys, and the review itself.
Policy and risk owners
A policy the product reads, not one it is told about
The classification matrix: what an agent of each class, holding a tool of each tier, may do. Versioned rather than edited, so last month’s permissions stay readable.
Security and architecture
The tool nobody approved
One seat says what exists; a different seat says what it is trusted with; and an approval is granted against the exact wording somebody read.
Security teams
Where the data goes, and who holds the keys
Providers and the reference each key is held under, the dataset catalogue, and the one thing no screen will ever report: what an individual person did with an agent.
Assurance and quality
When the review has quietly become a stamp
Verdicts recorded against a rubric version, reviewer agreement and seeded errors to check the checkers, and a maturity level computed from that evidence rather than claimed.
Looking at the whole estate
For the people who have to answer for all of it at once.
Risk and assurance leads
Five scores, and why they are never added together
Responsibility, auditability, interpretability, dependability, traceability — each scored from evidence that was recorded, and each capped by the evidence that never appeared. A single number would hide the one that is failing.
Boards and programme sponsors
A maturity level you cannot talk your way into
Walked from the register rather than answered in a questionnaire. The ladder stops at the first rung with a blocker and names it, and no field anywhere lets somebody assert a level.
Department and practice heads
The same report, for your department only
Agent, team, practice, estate: one ladder, and your seat sits on one rung of it. A request above that rung is refused before the database has been asked anything at all.
Risk owners
A risk register ordered by what to do on Monday
Weaknesses grouped by the one fix that would clear them, most first, each owned by a seat rather than a person — and the ones nobody has grouped yet keep their own heading instead of being swept away.
What it costs, and what you can prove about that
For whoever signs the invoice, and whoever is asked to justify it.
Finance and platform owners
Any tool can show you a number. Ask where it came from
A run the provider did not price holds no cost at all, and is never counted as free. The total arrives with the share that is missing from it, and a figure covering measured and assumed prices is reported as assumed.
Boards and budget holders
Cost per successful task, not cost per attempt
Divide by every attempt and a failing agent looks cheap. The denominator is named on the page rather than assumed, and the figure carries its provenance.
Programme and finance leads
The bill nobody forecast, and the two things behind it
One agent, one quarter, spend up a hundredfold. Volume and unit price are kept apart, because only one of them is a problem and a single figure cannot tell you which.
Finance business partners
Which department to charge, and the row nobody else shows
Practice, team and agent, each row with its own provenance — and unallocated spend named rather than spread across the others, which is the usual trick and makes every other row wrong.
Platform and risk owners
A spend limit that has never fired has never been tested
Runs stopped by a limit are counted and shown as the control working. Colour them red and everybody learns to turn the limits off.
Sponsors and sceptics
Does the governance cost more than it saves?
The counts are here: work put in front of a person, work refused, runs a limit stopped. The money cost of review is not, because no rate for it is recorded and a saving needs a version of events nobody observed.
About these recordings
Recorded on the running product. Any figure spoken was read off the page on screen, which is the rule the product holds itself to. The estate is our own and is there for demonstration: it is not a customer’s, and no customer’s data appears in any frame. The voice is synthetic.
See it working
Thirty minutes, on screen. We will show you what runs today, say what is still being built, and answer what we cannot do yet.
Book a working demonstration