Every AI turn has to run somewhere. Argand decides where.
Most agent work doesn’t need a frontier model — but almost all of it gets one, because nothing is deciding otherwise. Argand starts every turn at the cheapest rung that can hold it, and moves up only when the work says so.
Not an IDE. Nobody sits in it.
It sits underneath the tools you already use.
Argand is an operator‑governed, algorithmically agentic, self‑calibrating, distributed network protocol for carrying state across any runtime and interface.
The protocol has a name: AAP, the Argand Arbitration Protocol. Argand is the product; AAP is what it speaks. A client someone else writes implements AAP — not Argand — which is the whole difference between a layer and an app.
The agent — model plus harness — is selected per task rather than configured. Everything on the real axis gets cheaper every quarter; nothing on the imaginary one does, which is why it’s the half we don’t automate.
That sentence describes where this is going, not where it is. See Where it stands.
What gets decided, every turn
Four of them optimise downward. Two optimise the other way — and those two are what make it something a stranger can run.
Runtime
Which machine, whose machine, and which server turns weights into tokens.
laptop · your serverrented GPU · API
Model
Frontier when the work needs it. Open weights when it doesn’t — which is most of the time.
frontier APIopen weights
Harness
Which agent loop wraps the turn. The axis nobody else automates.
Claude Code · CodexGemini · raw API
State
What the turn resumes from — the workspace, the loaded context, the topic. Carried between machines, not pinned to one.
workspace · contextAuthority
What this turn is allowed to do. Read, draft, or reach the outside world — enforced by the machine, not by convention.
Not built yetBudget
What it may spend, checked before anything is provisioned rather than after the bill arrives.
Not built yetBottom rung first
Not “which model is best here” — that’s a prediction problem nobody has solved. Just “is the cheap one good enough,” which is a threshold, and thresholds are answerable.
Starting at the bottom is nearly free to get wrong. A failed attempt on hardware you already own costs seconds, not dollars — so the question isn’t how often the cheap rung is right, it’s how fast it fails when it isn’t.
And nothing gets offloaded that can’t be checked. A test, a schema, a diff, a second pass. The checker is what triggers the escalation — which means the savings are bounded by what’s checkable, not by what’s cheap. We’d rather say that than discover it later.
Most agent work is cheap work
Extraction, drafting, classification, routing, screening. None of it needs the most capable model available, and all of it currently gets one.
At any plausible seat price, routing every turn to the most capable model costs more than the seat. This is the wall every agent platform is currently walking into.
The default, todayAn independent August 2026 study answered more questions correctly than the best single model, at $3.33 against $7.69 — by assigning models per task type rather than picking one. Small study. Same direction as Stanford’s FrugalGPT.
arXiv:2608.23023 · not oursOthers have built the parts. Nobody routes.
An earlier version of this page claimed the market had split in two and neither side crossed. That turned out to be wrong, and we’d rather correct it here than defend it.
| Agents | Runtime | State | Routes it | |
|---|---|---|---|---|
| Model routers | ✓ | — | — | — |
| Compute schedulers | — | ✓ | — | — |
| Multi-harness clients | ✓ | partial | ✓ | — |
| Hosted agent platforms | ✓ | one cloud | ✓ | — |
| Argand | ✓ | ✓ | ✓ | ✓ |
Every product in that table is a picker. You choose the harness, you choose the machine, you choose the model — by hand, every session. Good tools, and not one of them decides for you. That last column is the entire thesis, and it’s the only one we’d defend.
Run it yourself, or don’t
Same arbitration layer at every rung. What changes is who owns the machine and who carries the operations.
Honestly, early
The execution layer runs in production today — on one operator’s system, which is the point of saying so. Everything below is measured or marked unbuilt; nothing is projected.
- Per-turn execution target across state, tier, harness and model
- Four agent harnesses behind one interface
- Cost accounting by tier
- Records where a turn actually ran, not just where it was sent
- Approvals that fail closed, and admission control
- Rented-GPU provisioning
- The spend ceiling — and nothing gets rented until it exists
- Machine-enforced authority per turn
- State that survives a hop between machines
- Automatic selection — the part that matters most
It’s being built in the open.
No waitlist theatre and no launch date. If you run agent work at enough volume that the bill has started to matter, that’s the conversation worth having now — while the routing policy is still being shaped by real workloads.
Alpha. Invite only. Nothing has been validated on anyone but the operator — and that’s the first thing we’re trying to fix.