AI coding on a fraction of the energy.
It starts with energy observability — Agent Joule meters every edit in real joules. Then its deterministic-first cascade resolves most changes locally, for picojoules, and the energy of writing code falls exponentially.
A session, metered
Most edits never reach a model.
One real change — refactor an auth module — closing tier by tier, each op priced in joules as it lands.
The thesis
AI coding can usher in energy-optimized compute.
Today, each AI edit is a datacenter round-trip — energy per change is rising. It doesn't have to. Three moves turn the curve around, and they build on each other.
Energy observability
You can't optimize what you can't see. Agent Joule meters every edit in real joules, read from the host's own power counter — the status line shows exactly what a change cost.
Energy-optimized compute
Seeing the cost lets you drive it down. A deterministic-first cascade resolves most edits with lookup, search, and templates — for picojoules. A model is the last resort, not the default.
Exponential decrease
The cascade learns. Closed work is cached and retrieved, so a growing share of edits resolve at the cheapest tier — and the energy of writing code falls, edit over edit.
01 · Energy observability
You can't optimize what you can't measure.
Agent Joule reads energy from the host's own counter wherever the hardware exposes one — the same probe, the same way, across the ecosystems developers use. Every change carries its real cost in joules. That measurement is the foundation; nothing else is possible without it.
| Hardware | Counter | Provenance | Reading |
|---|---|---|---|
| Apple Silicon | IOReport Energy Model | ModelBased | 25.5 W / 74 J |
| Windows + RTX 2060 | NVML | HwShunt | 9.4 W / 18.8 J |
| Linux x86 + Tesla T4 | NVML | HwShunt | 71.8 W / 148 J |
| Linux x86 Xeon | RAPL powercap | HwShunt | 77.1 W / 154 J |
Verified on real silicon. Each reading declares its provenance — HwShunt for a hardware counter, ModelBased for a vendor energy model — so the number is honest about how it was known.
02 · Energy-optimized compute
The cheapest method that can close the change, wins.
Once you can see the cost, you can drive it down. JouleClaw runs every request down a cascade and stops at the first tier that resolves it. Lookup, search, and templates — deterministic, near-free — take precedence. A model is the last thing tried, not the first.
The change was resolved before. A content-addressed lookup returns it instantly, for picojoules — exact, reproducible, free.
Answerable from a search over your tree and prior changes — the nearest known solution, surfaced without generating anything new.
A formatter, a codemod, a rename, a scaffold — a transformation with one correct answer, applied by rule. No model involved.
When generation is genuinely needed, a local model runs first — on your hardware, metered in joules, nothing leaving the machine.
When nothing else can close, a model is invoked — free-tier, paid, or openweight, your own key. The rigid harness elevates whichever you bring. On most changes, it's never reached.
03 · The exponential
The more you code, the cheaper each edit gets.
The cascade has a memory. Every closed change is content-addressed, cached, and made retrievable — so a repeat resolves at L0 for zero joules and zero dollars, 140× faster. As that hit-rate climbs, the average energy per edit decays toward the deterministic floor.
Proof · joule-bench
Same resolution. ~10× lower cost and energy.
On a realistic operation mix, the cascade reaches the same resolution as a frontier cloud agent — because it escalates novel work to the same model at L4 — while most edits close locally and never make the trip. Measured by an executable test oracle paired with host-counter energy and vendor billing.
Accounted for
Every edit is an object you can name, sign, and price.
Observability isn't only energy — it's provenance. Each change is a content-addressed, signed coding operation with a receipt: what changed, on what base, at which tier, for how many joules. No phone home; the trail stays on your machine.
- op
- b3:Ynpl53wU4pwfokr8h2c…
- prior
- b3:8fK2q9Lm4nRtZ0xV…
- tier
- L1 · Template
- energy
- 0.004 J · HwShunt
- sig
- ed25519 ✓ verified
Content-addressed
Every change is named by its BLAKE3 hash over a canonical encoding. The same change always has the same id — a stable fingerprint for every edit.
Signed & attributed
Each operation is an Ed25519-signed transition bound to the prior state of the tree. Who changed what, on what base, by what tier — provable, not assumed.
Receipted in joules
Each change carries a receipt: the tier that produced it and the joules it cost, with honest provenance — lookup or frontier model, you can see which did the work.
Reproducible
The deterministic tiers give the same output for the same input, every time. Most of your history is replayable, not a random walk.
Private & federated
Nothing phones home. Optionally, a fleet agrees.
The loop runs on your machine — no sign-in, no telemetry, no auto-update, with a build-level egress backstop. When you want scale, nodes federate over QUIC and reach byte-identical agreement by BFT consensus: a result reproducible across a fleet, none of them trusting a server.
The pillars
Built on the open standards of Transaction Science.
The deterministic discipline, the energy meter, and the signed trail aren't bolted on — they're open standards with real reference implementations, shared across the family.
JouleClaw
The master agent. It owns the loop — the deterministic-first cascade, the on-device runtime, the energy meter — and wraps every model call in a rigid harness that elevates whatever you bring.
Joule Code
The open standard the cascade speaks: content-addressed, signed coding operations, the L0–L4 tiers, and JCR-1 receipts. Agent Joule is its reference consumer.
Transaction Science
The family the pillars come from — open standards that make digital work content-addressed, signed, metered, and conformance-defined.
Get it
Install Agent Joule. Watch the meter.
A fork of Zed, in the open — Apache-2.0 lineage, pinned to the Transaction Science pillars by git revision. Install it, point it at any model key, and let JouleClaw reach for the deterministic answer first.
Picks your chip automatically. Direct .dmg — Apple Silicon → aarch64, Intel → x86_64 — from Releases.
x86_64 & arm64, after adding the apt repo. Or .deb / .tar.gz.
x86_64 & arm64, after scoop bucket add transaction-science …. Or the .exe / .zip from Releases.
First launch on macOS
The build is open-source and unsigned (no Apple Developer toll), so Gatekeeper warns once on first open. Homebrew installs clean — for a direct .dmg, run this one line, then open normally:
Or: right-click the app → Open → Open. On recent macOS: System Settings → Privacy & Security → Open Anyway.