Huxley–Gödel Machine · AGI Jobs

Turn intelligence into useful work, and use reviewed outcomes to guide the next investment.

Open the research console · Operator runbook · Validation and limits

AGI Jobs is designed as a scalable machine labor layer for authorized, lawful screen-based work—coordinating specialized agents to execute tasks, produce reviewable evidence, and support independent verification and settlement across a broad range of computer-based workflows.

This module makes the improvement loop inspectable: compare hierarchical exploration with a greedy baseline, enforce an experiment budget, inspect the lineage, and produce a source-bound benchmark analysis. The long-term ambition is a productive network that can reinvest demonstrated value into better research, software, infrastructure and eventually greater energy and industrial capacity. Scale follows evidence, resource availability and governance.

Start in two minutes

From the repository root, with Python 3.12:

python demo/Huxley-Godel-Machine-v0/run_demo.py --seed 7

The standard-library simulator writes reports to demo/Huxley-Godel-Machine-v0/reports/ and its current comparison to demo/Huxley-Godel-Machine-v0/web/artifacts/comparison.json. Open the published console and import that comparison file. Uploaded files are processed in your browser; the page does not dispatch workers, connect wallets or send payments.

For a local, self-contained viewer, use the repository's pinned Node/npm versions (.nvmrc, package.json) and locked dependencies:

npm ci
npm run demo:hgm:build
python -m http.server 8765 --bind 127.0.0.1 --directory build/hgm

Open http://127.0.0.1:8765. The build includes three freshly generated recordings: reference, constrained budget and owner pause. Import a new comparison to view your own run; the bundled reference does not update when another process runs. Stop the local viewer with Ctrl+C.

What is implemented

The scope spans software, research, public datasets, web applications, scientific reproduction, AI evaluations, editable documents/presentations, vendor research, tooling and other authorized computer workflows. Tool availability expands what can be attempted; it does not prove that every task can be completed reliably. Use approved public, licensed or synthetic non-personal inputs. This module's connected benchmark task uses synthetic input only.

Read the numbers correctly

• gmv is simulated gross value, in USD-equivalent units. It is not a USDC balance, customer revenue or settled work.
• The historical roi field is gross value / completed cost, a gross multiple. Net return would be (gmv - cost) / cost. Undefined ratios serialize as JSON null.
• reserved_cost is the cost of queued, unfinished work. Admission checks cost + reserved_cost + next_cost <= max_budget before scheduling. Reports retain these commitments at the horizon instead of pretending they completed.
• pending_tasks records unfinished work. The console's value-less-commitments metric is gmv - cost - reserved_cost.
• Both strategies observe the same budget ceiling and owner directives. Compare seeds, evaluation counts, elapsed steps and costs; a favorable single seed does not establish superiority.
• The website's $40 trillion/year opportunity is a user-supplied planning assumption. The calculator applies eligibility and capture percentages; its result is illustrative gross work value before costs, not measured TAM, platform revenue or a forecast.

Run a bounded experiment

python demo/Huxley-Godel-Machine-v0/run_demo.py --seed 7 \
  --set economics.max_budget=100 \
  --output-dir /tmp/hgm-budget-study \
  --ui-artifact /tmp/hgm-budget-study/comparison.json

python demo/Huxley-Godel-Machine-v0/run_demo.py \
  --set owner_controls.pause_all=true \
  --output-dir /tmp/hgm-paused \
  --ui-artifact /tmp/hgm-paused/comparison.json

Malformed values, non-finite numbers, unknown configuration keys, negative costs and inconsistent ranges fail before output creation. Owner pause prevents new scheduling; it does not revoke effects already dispatched in another system. The simulator is a finite experiment, not a continuously running controller.

Outputs: effective_config.json, summary.json, summary.txt, hgm_timeline.json, baseline_timeline.json, hgm_lineage.mmd, roi_comparison.svg, logs.md, plus the requested comparison file. Save different experiments to different directories. The same seed and resolved configuration reproduce metrics; timestamps and destination paths vary.

Preserve the system map

flowchart LR
    Operators((Mission Owners)) --> demo_Huxley_Godel_Machine_v0[[Demo → Huxley Godel Machine v0]]
    demo_Huxley_Godel_Machine_v0 --> Core[[AGI Jobs v0 (v2) Core Intelligence]]
    Core --> Observability[[Unified CI / CD & Observability]]
    Core --> Governance[[Owner Control Plane]]

This is the original conceptual systems map. The static console is an observer and authoring surface; integration with live orchestration requires the explicit admission process in the runbook. The original web evolution flow and Grand Operator Console diagrams are also retained. Their terms “self-modification” and “mission execution” describe simulated quality mutations and sampled outcomes here, not changes to production code.

Directory and compatibility guide

make demo-hgm and python -m demo.huxley_godel_machine_v0.simulator remain supported. The historical run.py path remains available for comparison. Do not mix configuration or telemetry formats between implementations.

Verify before merging

PYTEST_DISABLE_PLUGIN_AUTOLOAD=1 python -m pytest demo/Huxley-Godel-Machine-v0/tests -q
npm run demo:hgm:build
npm run demo:hgm:test
npm run demo:hgm:lint
npm run demo:hgm:qa

Changes land through a pull request with required checks green. The dedicated workflow exercises the simulator, task inspector and browser; the Pages workflow also checks the published HGM route. No live deployment, unrelated reviewer acceptance, buyer use, payment or production commissioning is asserted by these checks.