ai-agents-research-frontier
Three ranked open research programs for this repo, each with honest current-state evidence, first concrete steps, and a falsifiable milestone. Verified governance (ADR-069, proposed), cross-harness abstraction (ADR-072 proposed, ADR-068 accepted), and the self-improving loop (issue #1345). Use when you say `research frontier`, `open problems`, `what should we research next`. Do NOT use for how to run an experiment here (use `ai-agents-research-methodology`).
rjmurillo · 已有静态检查记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add rjmurillo/ai-agents --skill ai-agents-research-frontier技能说明
name: ai-agents-research-frontier
description: "Three ranked open research programs for this repo, each with honest current-state evidence, first concrete steps, and a falsifiable milestone. Verified governance (ADR-069, proposed), cross-harness abstraction (ADR-072 proposed, ADR-068 accepted), and the self-improving loop (issue #1345). Use when you say research frontier, open problems, what should we research next. Do NOT use for how to run an experiment here (use ai-agents-research-methodology)."
version: 1.0.0
license: MIT
metadata:
routing:
role: conditional-adjunct
invoker: ai-agents-research-methodology
trigger: ai-agents-research-methodology points here for the open research programs
user-facing: false
AI Agents Research Frontier
Open problems where this repository can advance the state of the art, ranked in
owner-confirmed priority order. This skill tells you WHAT is worth working on and
what "done" would look like. It does not teach experiment discipline (that is
ai-agents-research-methodology) and it is not the portability battle plan
(that is ai-agents-portability-campaign).
Honesty contract for this document: every asset claim below was re-verified against the working tree on 2026-07-30. Anything labeled PROPOSED is not policy. Anything labeled UNVERIFIED could not be confirmed and must be checked before you build on it.
Triggers
research frontieropen problemswhat should we research nextfrontier programs
The Three Programs at a Glance
| Rank | Program | Anchor artifacts | Status (as of 2026-07-30) | Falsifiable milestone (short form) |
|---|---|---|---|---|
| 1 | Verified governance | ADR-069, scripts/eval/eval-rule-activation.py | ADR-069 is PROPOSED; eval tool exists, 15 rule scenario fixtures exist | Controlled eval shows gated-corpus sessions beat ungated on N scenarios with defensible stats |
| 2 | Cross-harness abstraction | ADR-072, ADR-068, build/generate_agents.py, tests/build_scripts/test_generate_hooks_runtime_contract.py | ADR-072 PROPOSED, ADR-068 ACCEPTED as of 2026-07-30; generators and contract tests exist and run | New harness target added with zero hand-edits to generated trees, contract suite green |
| 3 | Self-improving loop | issue #1345 hooks, EVENT= telemetry (retired, ADR-084/#5154) | Apply-step hooks unregistered; end-to-end consumer pipeline UNVERIFIED | A correction observed once auto-proposes a guard that survives calibration |
Process
Phase 1: Verify current state before claiming anything
The evidence below decays. Re-run these before quoting any of it in an ADR, issue, or write-up. All commands run from the repo root.
| Claim | Re-verify command |
|---|---|
| ADR-069 status is proposed | head -5 .project-toolkit/architecture/ADR-069-context-corpus-is-the-product.md |
| ADR-072 status is Proposed with approval conditions | sed -n '1,15p' .project-toolkit/architecture/ADR-072-jtbd-plugin-architecture.md |
| ADR-068 status is Accepted as of 2026-07-30 | sed -n '1,10p' .project-toolkit/architecture/ADR-068-consolidated-hook-dispatcher.md |
| Rule-activation eval exists with a no-spend path | python3 scripts/eval/eval-rule-activation.py --help |
| Rule scenario fixtures | set -- tests/evals/rule-scenarios/*; echo $# |
| Corpus size across skills, rules, retros, memories | python3 -c "from pathlib import Path as P; print(len(list(P('.claude/skills').glob('*/SKILL.md'))),'skills',len(list(P('.claude/rules').glob('*.md'))),'rules',sum(1 for p in P('.project-toolkit/retrospective').glob('*.md') if p.is_file() and p.name != 'INDEX.md'),'retros',sum(1 for p in P('.serena/memories').rglob('*.md') if p.is_file()),'memories')" |
| Runtime contract tests pass | uv run pytest tests/build_scripts/test_generate_hooks_runtime_contract.py -q |
| Apply-step hooks unregistered | uv run pytest -q "tests/build_scripts/test_copilot_dispatcher_artifact.py::TestDispatcherArtifacts::test_retired_hooks_are_absent_and_keepers_are_plugin_only" |
Phase 2: Pick a program
Take the highest-ranked program whose first unclaimed step you can finish. Rank
order is owner-confirmed; do not reorder it without the owner. If your idea does
not fit any program, it is a new capability: route through the
buy-vs-build-framework Quick tier gate first (AGENTS.md Skill-First section),
then ai-agents-research-methodology for the idea lifecycle.
Phase 3: Execute the first steps
Each program section below lists three concrete first steps IN THIS REPO. Steps
are ordered; do not skip to step 3 because it is more interesting. Every step
that produces a measurement must predict the number before running (see
ai-agents-research-methodology), and every probe of external tool behavior
must follow ai-agents-empirical-probe-toolkit (pinned version, foreign
cwd/env, negative control).
Phase 4: Report and gate
Results go through the normal machinery, not around it: write-ups per
ai-agents-docs-of-record, changes through ai-agents-change-control, ADR
edits fire the adr-review debate gate. A negative result is a result: record
it (the #2230 rejected-fix pattern) so the next person does not re-run it.
Program 1: Verified Governance (ADR-069, PROPOSED)
Thesis (quoting the ADR title verbatim): "The Curated Context Corpus IS the
Product, Orchestration Is Plumbing"
(.project-toolkit/architecture/ADR-069-context-corpus-is-the-product.md, status:
proposed, date 2026-05-02). Core frame: no learning between runs; the only thing
that persists is what re-enters the next context window, so the corpus is the
durable competiti