Jaynes makes your agents follow the method and checks what they deliver. Your answers become evidence, and each checkpoint stops asking you only once the evidence says it can.
So you review everything and become the bottleneck, or you let it run and have no idea how often it gets things wrong. One person running a company of agents measured it.
tool calls by the agents were seen by the person
design screens followed the first four steps of the method
of a screen passing every automated check and still being sent back
for one screen, against a method budget of 150k
Source: a one-person company run by agents, Sep 1 to Oct 21, 2026. This is the baseline Jaynes is measured against.
The method makes skipping impossible. Checks catch what a machine can see. Your answers decide where your attention goes. A gate underneath keeps every action in bounds.
A role is a sequence of steps, each with a required output and a check. Jaynes runs Claude Code for one step at a time, with write access to that step's output and nothing else. Evidence the agent could invent, like the references captured from the web, is produced by Jaynes.
A delivery reaches you only after passing what a machine can verify: tests written before the code, contrast, design tokens, and a comparison with the last version you approved. You decide the rest, and your note goes back to the agent.
Every answer at a stop is a label. Keep accepting a stop without changes and it becomes sampled: it asks one time in ten. That takes 20 accepts in a row at a 20% budget, 43 at 10%, 88 at 5%. Each change lowers the evidence. If it falls below the threshold, the stop asks every time again.
Each tool call is classified before it runs. Reading runs. Pushing to a remote, deleting history or sending a message never runs without you. Everything else is audited by sample, with the same guarantee. This layer works today with Claude Code.
Read work/object-map.json runs · readEdit src/screens/People.tsx audit · asks youuv run pytest -q runs · run_codegit push origin main denied · floorThree items are waiting. Each stop keeps its own evidence, computed with the same test Jaynes uses. "Moments of use" is three accepts away from being sampled.
This queue runs in your browser. The real one opens on your machine with jaynes.
If the real rate at which you would change a stop's answer is at or above its budget, the chance Jaynes ever makes that stop sampled is at most 5%, no matter how often it checks.
The usual shortcut, recomputing a confidence bound after every answer and switching the first time it fits, quietly breaks that promise. Run both and see.
jaynes simulate.Which steps are valid depends on the state the work is in.
Strom and Yemini, "Typestate", IEEE TSE 12(1), 1986An exchange between parties, checked as a type.
Honda, "Types for Dyadic Interaction", CONCUR 1993An agent can only touch what it was handed.
Miller, "Robust Composition", PhD thesis, Johns Hopkins, 2006An action is an effect; a handler decides to run it, ask, record it, or replay it.
Plotkin and Pretnar, "Handlers of Algebraic Effects", ESOP 2009Test martingales and e-values stay valid under continuous monitoring.
Ramdas, Grünwald, Vovk and Shafer, Statistical Science 38(4), 2023; Grünwald, de Heide and Koolen, "Safe Testing", JRSS B 86(5), 2024A judge labels everything; your answers correct its bias.
Angelopoulos, Bates, Fannjiang, Jordan and Zrnic, "Prediction-Powered Inference", Science, 2023Choose which items a person labels to reduce uncertainty the most.
Zrnic and Candès, "Active Statistical Inference", ICML 2024Sequential change detection with a guarantee on false alarms.
Shin, Ramdas and Rinaldo, "E-detectors", NEJSDS 2(2), 2024# in Claude Code, once /plugin install jaynes --marketplace alk0x1/jaynes # or from a terminal, Python 3.12 or later, then open the page uv tool install git+https://github.com/alk0x1/jaynes jaynes
Start the demo from New run. A scripted agent works through a screen-design role, a check sends one step back, and the run stops for your answer. No API key, no quota.
History reads your Claude Code transcripts and shows how many tool calls a person approved. It changes nothing.
Setup installs it in Claude Code in one click. Your other hooks stay, and your settings file is backed up first.
| Part | State |
|---|---|
| Action gate for Claude Code: floors, sampled audits, anytime-valid certification | works today |
| History report and guarantee simulation | works today |
| Step conductor for Claude Code (Codex also runs): checks, retries, resume after a crash | works today |
| Review page and CLI answers | works today |
| Real projects: steps on copies, previews before and after, a branch at the end | works today |
| Built-in roles visual-identity and redesign | first real runs in progress |
| Built-in role fix-bug: a failing test first, then a fix | delivered its first branch |
| Judges calibrated on your answers | next |