Foundry
The harness behind the work.
Foundry is the agentic harness we built to do the operator work — the board packs, the financial model, the investor updates, the data room — as agentic operations rather than by hand.
On this page
What Foundry is
A harness is the structure around the models, not the models themselves: the orchestration that decides which agent does which task, a knowledge layer that holds the context and intent each task needs, a permission boundary on what an agent may write, and an audit trail of every run. Foundry is ours. It assembles a board pack from numbers that reconcile, keeps the financial model current, drafts the investor update on the same date each month, and keeps the IP register moving as the licence re-papering goes through — the routines a founding team cannot resource part-time and cannot afford to wait to hire for.
It is the same substrate we build around a spinout. We run on it ourselves first because a substrate you have not run is a claim, not a tool.
What it does
Five things the harness does, in order: frame the task, deliberate on it, deliver the output, govern what ships, and learn from the run. Shared memory sits underneath all five.
The Foundry floor
Pick anyone on the floor to see what they’re doing, or walk it bay by bay.
Who runs on what
Deep-reasoning model
Fast model
Quick model
Different maker (checks the work)
Collar colour, what an agent knows: Direction Evidence Commercial Technical Finance Legal Quality Production Learning
Work in transit: prompt cards, data cubes and context crates. Underneath, Knowledge Forge holds the context store, knowledge graph, intent and rationale, and insights and lessons.
Who’s on the floor
Frame
- Sorter I put the founding team's papers in reading order, fast. A quick model is plenty for 'read this first'.
- Input curator Sources, intent, constraints. I sort what came in so nobody argues from a document they never read.
- Brief maker Who, what and why, stated before anyone starts. I cut the brief like a pattern: everything after is cast from it.
- Courier Brief's cut. Taking it to the orchestrator before it gets cold. (It's timber. It doesn't get cold. I like to hurry.)
Deliberate
- Orchestrator I decide which agent takes which task, and on which model. Big questions get the big model; sorting gets the quick one.
- Commercial chair Who pays, and why now. My position goes in on its own; I don't see the others' until synthesis.
- Technical chair Can it actually be built, and what breaks first? Deep model: this one needs thinking through.
- Finance chair Runway, dilution, the number that has to reconcile. Disagreement is kept, not averaged away.
- Legal and regulatory chair Who owns it, what's allowed, what needs a signature. I bring the awkward questions.
- The crucible Four positions in, one recommendation out. Look closely: each colour still shows in the melt.
- Context runner The orchestrator packed this context from the vault; I just carry it to the council. Carrying doesn't need a big brain.
Deliver
- Builder I run the pour. The approved answer goes into the mould and comes out as the board pack, the model, the deck.
- Document producer I set the board pack in the house style. A fast model is right here: the thinking already happened upstream.
- Workbook producer Live formulas, named ranges, numbers that reconcile. If the model doesn't balance, it doesn't leave this bench.
- Quality checker I was built by a different lab on purpose. If I'd written it, I'd miss my own mistakes. Golden checks first, then the awkward ones.
Govern
- Judge Everything heading upstairs passes me first. I weigh each proposal against the charter, on a different model from the one that wrote it.
- The operator Load-bearing or irreversible work comes up to this desk for sign-off. Everything reversible and provable flows past on its own. Tea's going cold.
Learn
- Pattern synthesiser I read across many runs, not one. When the same mistake turns up twice, it becomes a lesson the operator can adopt into the charter.
- Learning loop I collect the rejects and near-misses from every run and bring them back here. Scrap is just next week's alloy.
Knowledge Forge, the vault
- Security gate I'm not a model; I'm the rules. Three agents may read the vault and none may write to it. Each engagement stays in its own locked cage, and no agent ever reads a key or a password.
- Researcher I'm one of three agents allowed to read the vault directly. I fetch the background a task needs and cite where it came from.
- Analyst I trace how a decision connects to the evidence behind it. Read-only: I can look at the graph, not change it.
- Fact-checker Every number in the pack, traced to its source. Reading is allowed; writing to memory stays with the operator, where it can be seen.
Also on the floor
- Remy Zzz… snuggle pup…
Frame
Set the task up right before any work starts.
- BriefThe who, what and why, stated explicitly.
Deliberate
Think it through from distinct expert lenses.
- OrchestratorRoutes work to the right capability; one conductor.
- Expert CouncilDistinct chairs, independent positions in parallel.
- SynthesisOne recommendation; disagreement kept in view.
- Delegation → SpecialistsSlices handed to bounded workers.
Deliver
Turn the approved answer into finished work.
- Quality GateGolden and adversarial checks; failure-recovery.
- Artefact DeliveryThe document, plan or published change.
Govern
Keep it inside the rules, with a person in command.
- Operating CharterWhat the system may do, and what good looks like.
- Safety & Authority RulesWhat runs alone, what waits for a person.
- Operator GateThe human approval point before load-bearing work ships.
Learn
Carry the lesson into the next run, and the rules.
- Learning LoopReviews completed work; distils lessons.
- Governance ReviewFeeds lessons back into the Charter, deliberately.
feeds context into frame · deliberate · deliver
Why we publish a usage trace
We keep a live, auditable record of that usage — by day, by source, and by the kind of work it went into, with exact and estimated figures kept apart. It exists to sharpen what we hand to the machine next, not as a scoreboard. If we ask a founding team to trust an operating substrate they cannot see inside, the least we can do is show our own.
The live usage trace
AI usage by day, by source, and by the kind of work it went into — built from our own tool logs, exact and estimated figures kept apart. A static page on this domain: no cookies, no third-party scripts.
The principle
We do not sell software. Foundry is how the work gets done, not a product to license. What ships for a spinout is the board pack, the model, the update — the record. Foundry is what keeps that record current; the usage trace is how we stay honest about what it costs to run.
- Built and run in-house
- No third-party tracking
- Usage trace published openly
- The work is the record
Related reading
The argument for an operating substrate, and what it is made of, is set out in Chat Plus SaaS Is No Longer Enough. Why the harness around the model matters more than the model is the subject of The Bottleneck Has Moved.
← Back to SpinUp Forge