Agentic Principles
Research for AI agents that do real work.
This project asks how agents can plan, use tools, and change external state without outrunning their authority, evidence, or ability to recover. It studies software delivery, automatic SRE, customer service, and other consequential SaaS operations through source research, live tests, transcript analysis, and evaluations.
The result is not a list of commandments. It is a reviewable catalog of claims, each labelled by the strength of its evidence and linked to the work behind it.
Start with the reader’s guide, or scan the current principle catalog.
Current evidence posture
- 1 candidate principle has a scoped mechanism, counter-pressure, falsifier, and mixed-method study: contain a partial failure and continue only the independently verifiable safe frontier.
- 10 seeds are early research directions, not recommendations.
- 0 supported principles have yet met the independent empirical promotion standard.
The partial-failure study is the strongest current result and the best example of the intended method.
Repository map
| Path | Purpose |
|---|---|
docs/index.mdx | Human-oriented entry point and guide to the evidence labels |
docs/VISION.md | Research mission, scientific method, and intended product feedback loop |
docs/research/ | Timestamped questions, methods, evidence, analyses, and limitations |
docs/principles.json | Machine-readable principle registry |
.engineering/schemas/ | Project-owned JSON Schema contracts |
.engineering/planning/ | Governed planning artifacts, mutated only through aep artifact |
website/ | Docusaurus source for the public research site |
More columns: swipe horizontally, or focus the table and use the arrow keys.
Principles move through an explicit lifecycle:
seed → hypothesis → candidate principle → supported principle
↘ challenged → revised or retired
The maturity label is a claim about evidence strength, not editorial polish. Read
docs/index.mdx for orientation, docs/VISION.md for the full
method, and AGENTS.md for the operating rules.
Verify the research contracts
The shared schema-contract tooling comes from
beyond10x/ess. From the
repository root:
ess schema validate \
docs/principles.json \
docs/research/evidence/scoped-progress-partial-failure/*.json \
--schemas .engineering/schemas
cd website
npm ci --ignore-scripts
npm run schema:check
npm run typecheck
npm run build
The authored JSON Schemas are the source of truth. TypeScript types under website/src/generated/
are deterministic projections and must pass the drift check.
Publication
Pushes to main that affect the research, schemas, or website run the GitHub Pages pipeline. It
validates schema instances and generated types, audits dependencies, type-checks, builds, and deploys
the static artifact with least-privilege Pages permissions.
node_modules, build, dist, coverage output, and Docusaurus caches are ignored and are never
release inputs. See CHANGELOG.md for released milestones.