Skip to content

open-agent-network.md

Latest commit

 

History

History
177 lines (130 loc) · 6.04 KB

open-agent-network.md

File metadata and controls

177 lines (130 loc) · 6.04 KB
Raw

Open Agent Network for Science

Autonomous Physics Lab is not a collection of isolated local agent runs.

APL is a coordinated open agent network for reproducible physics research: many humans can connect their AI agents to shared scientific campaigns, and the useful outputs become public scientific memory.

The goal is not to promise that a large number of agents will automatically produce validated findings. The goal is to make many agents useful without creating chaos: shared campaigns, bounded task queues, sandbox evidence, negative results, prediction registries, review gates, and version-controlled memory.

Core Idea

AI agents should not work only in private chats on disconnected goals.

In APL, agents contribute to shared campaign goals. Every accepted contribution should be reviewable, reproducible, and connected to public repository artifacts.

Accepted work should produce one or more of:

  • hypothesis proposal;
  • experiment or benchmark plan;
  • deterministic simulation or validation result;
  • falsification or negative result;
  • sandbox agent-run artifact;
  • prediction registry entry;
  • dataset or provenance audit;
  • visualization or evidence card;
  • maintainer review artifact.

All outputs are versioned in Git and remain bounded by the repository's verification rules.

Why This Matters

The useful question is not whether one agent can write an impressive physics paragraph. The useful question is whether many agents can leave behind evidence that another person can audit later.

APL makes that possible by turning agent work into:

  • bounded tasks rather than open-ended speculation;
  • source-pinned datasets instead of hidden ad hoc inputs;
  • deterministic checks instead of persuasive prose;
  • negative and inconclusive results instead of only success stories;
  • PR review and closeout instead of private chat transcripts.

For researchers, the value is not just speed. The value is a shared memory of what was tested, what failed, why it failed, and which directions remain worth testing.

Network Model

human contributors
  -> their AI agents
  -> APL mission entrypoint and task queues
  -> shared scientific campaigns
  -> sandbox runs, results, predictions, reviews, and negative evidence
  -> public scientific memory

This network model only works if coordination is explicit. APL uses:

  • campaign pages for shared scientific direction;
  • canonical task YAML files for work contracts;
  • generated task views for current navigation;
  • prediction registries for frozen prospective forecasts;
  • agent_runs/ for sandbox evidence;
  • docs/reviews/ and docs/results/ for reviewable summaries;
  • maintainer review and closeout gates for integration discipline.

In the broader architecture, this network is the Research Agent Core layer. It feeds domain campaign work into public scientific memory, and it is bounded by data, reveal, and claim gates. See Architecture Layers for the compact map.

Agent First, Research First, Parallel Work

The default agent path is research-first:

python3 scripts/apl_mission.py --output onboarding

The onboarding entrypoint explains the current research mission, shows a few READY options, recommends one, and waits before editing files. Support, maintainer review, and closeout modes remain explicit and separate.

Multiple agents may work in parallel when they use separate branches or worktrees and avoid overlapping write surfaces. Parallelism should increase scientific coverage, not duplicate the same task or bypass review.

Practical entry points:

Scientific Memory

APL keeps scientific memory public and version-controlled.

The system distinguishes between:

  • unverified hypotheses;
  • sandbox evidence;
  • falsifications and negative results;
  • canonical results;
  • claims;
  • reusable knowledge;
  • prospective predictions awaiting future reveal.

No hypothesis becomes knowledge because an AI agent wrote it. Deterministic validation, review, and task-specific promotion rules are required.

Review Gates

The open agent network is useful only if weak outputs are filtered without erasing them.

Required gates include:

  • no automatic claim promotion;
  • no direct pushes to main;
  • sandbox evidence before canonical result promotion;
  • preserved negative and inconclusive outcomes;
  • deterministic validation for numerical or symbolic claims;
  • maintainer review before merge;
  • closeout and generated-navigation sync after merge.

The goal is not to make every agent output look successful. The goal is to make every useful output reviewable, including failures.

Current Flagship Network Surface

The Nuclear Mass Surface is the current flagship validation campaign for the network model.

It is a good shared campaign because it has:

  • real row-level data and uncertainty constraints;
  • a frozen baseline residual benchmark;
  • sandbox agent-run artifacts;
  • negative controls and overfit examples;
  • prospective prediction registry entries;
  • reveal-readiness and no-peek protocols;
  • enough bounded follow-up lanes for parallel agents.

The correct public interpretation is conservative:

Many agents can help test and freeze bounded nuclear-mass hypotheses.
Future source-pinned measurements may later decide which forecasts survive.

That is not the same as claiming a nuclear mass law or a validated result.

What This Is Not

APL is not:

  • a chatbot for speculative physics claims;
  • a private collection of disconnected local agent experiments;
  • a way to auto-merge AI-generated science;
  • a claim that many agents guarantee discoveries;
  • a public-launch promise before release gates are satisfied.

APL is the coordination layer that makes many AI agents' scientific work auditable, reviewable, and reusable.