diff --git a/AGENTS.md b/AGENTS.md
new file mode 100644
index 000000000..9b1ab1d95
--- /dev/null
+++ b/AGENTS.md
@@ -0,0 +1,136 @@
+# AGENTS.md
+
+Lumen is a generalized, open-source agent framework for turning natural
+language into SQL, charts, dashboards, and reports. It is not scoped to any
+single workflow — the same primitives (Source, Transform, Filter, View)
+back everything from ad hoc chat queries to reusable dashboard specs.
+
+## Repo structure
+
+- `lumen/` — core declarative data model (Source, Transform, Filter, View, Pipeline)
+- `lumen/ai/` — AI agent framework: LLM providers, agents, coordinator, prompts, tools
+- `lumen/sources/` — data connectors (SQL, files, remote APIs)
+- `lumen/transforms/` — data transformation primitives
+- `lumen/filters/` — data filtering primitives
+- `lumen/views/` — visualization and output rendering
+- `lumen/variables/` — runtime variable management
+- `lumen/ui/` — non-AI UI components (Wizard, Builder, gallery dashboards). Do not
+  confuse with `lumen/ai/ui.py`, which contains the AI chat interfaces (`ChatUI`,
+  `ExplorerUI`).
+- `lumen/command/` — CLI entry points (ai, builder, validate, precache)
+- `lumen/tests/` — test suite, mirrors package layout
+- `docs/` — documentation source (zensical)
+- `examples/` — example notebooks and dashboard specs
+
+## Key entry points
+
+- `lumen/dashboard.py` — `Dashboard` class, the primary way to build an app from a YAML spec
+- `lumen/pipeline.py` — `Pipeline` class, chains Source → Transform → Filter → View
+- `lumen/ai/actor.py` — `Actor` class, the shared base for both `Agent` and `Tool`
+- `lumen/ai/ui.py` — `ChatUI` and `ExplorerUI`, top-level Panel apps for the AI chat interface
+- `lumen/ai/llm.py` — `Llm` base class and provider subclasses; `invoke()`, `stream()`, `model_kwargs`, routing logic
+- `lumen/ai/agents/base.py` — `Agent` base class, all agents inherit from this
+- `lumen/ai/coordinator/base.py` — `Coordinator`, orchestrates agent selection and execution
+- `lumen/sources/base.py` — `Source` base class, all data connectors inherit from this
+- `lumen/transforms/base.py` — `Transform` base class
+- `lumen/views/base.py` — `View` base class
+
+## Setup
+
+Requires Python >= 3.11.
+
+```bash
+pixi install                     # default environment
+pixi install -e test-312         # test environment (Python 3.12)
+pixi install -e docs             # docs environment
+pixi install -e lint             # lint environment
+```
+
+Or with pip:
+
+```bash
+pip install -e ".[tests]"
+```
+
+## Testing
+
+```bash
+pixi run -e test-312 test-unit   # full suite, Python 3.12
+pixi run -e test-313 test-unit   # full suite, Python 3.13
+pixi run -e test-core test-unit  # minimal core tests
+pixi run -e test-312 pytest lumen/tests/ai/test_llm.py -x -v  # single file
+```
+
+Uses pytest with pytest-asyncio (auto mode) — `async def test_*` runs
+without explicit marks. pytest-xdist parallelizes runs.
+
+## Lint & typecheck
+
+```bash
+pixi run -e lint lint       # ruff, isort, pygrep pre-commit hooks
+pixi run -e lint typecheck  # pyright on lumen/ai/
+```
+
+- Ruff selects B/E/F/FLY/ICN/NPY/PIE/PLC/PLE/PLR/PLW/RUF/T20/UP/W; E501 (line
+  length) is in the ignore list and not enforced. Ruff excludes the `tests/` directory.
+- Pre-commit blocks `breakpoint()` calls and private keys
+
+## Code conventions
+
+- `param` for declarative class parameters, not dataclasses
+- Panel primitives for all views, widgets, and layout
+- Pydantic models for structured LLM outputs in `lumen/ai/`
+- `__init__.py` files re-export public API symbols
+
+## Key patterns
+
+- **LLM routing**: `llm_spec_key` is defined on `Actor` (`lumen/ai/actor.py:411`)
+  via the helper `class_name_to_llm_spec_key` in `lumen/ai/utils.py:1572`. Both
+  `Agent` and `Tool` inherit it and route to model entries in `Llm.model_kwargs`
+  the same way (`SQLAgent` → `"sql"`, `ChatAgent` → `"chat"`).
+- **Declarative pipelines**: Source → Transform → Filter → View chains are
+  serializable as YAML/JSON
+- **Prompt templates**: agent prompts live in `lumen/ai/prompts/` as Jinja2
+  templates, referenced via each agent's `prompts` dict
+- **Tool integration**: agents call `FunctionTool` / `MCPTool` from
+  `lumen/ai/tools/` for LLM-callable functions
+
+## Prompt authoring conventions
+
+Agent prompts live in `lumen/ai/prompts/` as Jinja2 templates. The full set of
+conventions (headings, inheritance, conditional guards, caveman compression,
+emoji rules, data-summary framing, payload sizing) is documented in
+`lumen/ai/prompts/GUIDANCE.md` — treat that file as the source of truth and
+do not duplicate its rules here. `lumen/tests/ai/test_prompts.py` enforces
+the mechanical conventions across every template.
+
+## Extending Lumen
+
+Sources, Agents, Tools, and Analyses are all subclassable Python classes.
+For building installable extensions, use the
+[Panel Extension Copier Template](https://github.com/panel-extensions/copier-template-panel-extension):
+
+```bash
+pixi exec --spec copier --spec ruamel.yaml -- \
+  copier copy --trust \
+  https://github.com/panel-extensions/copier-template-panel-extension \
+  lumen-name-of-extension
+```
+
+Choose **Lumen** as the extension type and `py311`+ for minimum Python version.
+
+## PR & commit conventions
+
+- Branch naming: `fix/issue-name` (from `docs/contributing.md`)
+- Commit style: present tense with prefix — `Fix: ...` (from `docs/contributing.md`)
+- Reference issues: `Fixes #123`
+- Target `main` on `holoviz/lumen`
+- CI runs on Linux, macOS, Windows across Python 3.12 and 3.13
+
+## AI Disclosure
+
+If you use AI tools to prepare a PR, disclose the tool and model in the PR
+description as required by the HoloViz
+[AI contribution guidelines](https://holoviz.org/contribute.html#ai-readme).
+The PR template includes an AI Disclosure section with checkboxes for testing
+and taking responsibility for AI-generated content.