diff --git a/rfcs/0003-workflow-engines.md b/rfcs/0003-workflow-engines.md
new file mode 100644
index 00000000..480af6e6
--- /dev/null
+++ b/rfcs/0003-workflow-engines.md
@@ -0,0 +1,355 @@
+---
+rfc: 0003
+title: Delegate execution to workflow engines; deprecate recipes
+status: Draft # Draft | Active | Accepted | Rejected | Superseded
+authors:
+  - Pete Bachant (@petebachant)
+created: 2026-08-03
+tracking-issue: # link to the GitHub issue opened in Step 1
+superseded-by: # RFC number, once another RFC replaces this one
+---
+
+## Context
+
+Today, an ASTRA output says how it is built with a **`recipe`**: a `command`
+string, plus optional `container` and `resources`. A node-level
+`Analysis.container` supplies a default environment for the recipes beneath it.
+
+```yaml
+outputs:
+  - id: fit_params
+    type: table
+    inputs: [catalog_data]
+    decisions: [fit_method]
+    recipe:
+      command: >-
+        python src/fit_period_luminosity.py
+        --catalog {inputs.catalog_data}
+        --method {decisions.fit_method}
+        --out {output}
+```
+
+This is under-specified in two independent ways, and the second one is not
+fixable by adding fields.
+
+**A command does not describe its environment.** `python src/fit.py` reproduces
+only on a machine that already happens to have the right interpreter and the
+right packages. The `container` field was added to patch this, and it does help
+— but it means the spec now carries a *copy* of environment information that
+also exists wherever the project actually builds things, free to drift.
+
+**A command does not know whether it needs to run.** This is the deeper problem.
+A command always runs. It never notices that its input changed, and it never
+notices that it didn't. Nothing in a `recipe` lets a reader — or an agent —
+answer "is this artifact current?", which is the question that separates a
+reproducible result from a plausible-looking one. No additional field on
+`Recipe` fixes this, because staleness is a property of a dependency graph, not
+of a command string.
+
+Both failures land hardest on the case ASTRA is built for. An agent asked to
+produce an analysis will emit `python src/train.py` and consider the job
+finished. The document looks complete, validates cleanly, and cannot be
+rebuilt by anyone else. That is precisely the outcome the format exists to
+prevent.
+
+**Prior art solves this, and has for years.** Snakemake, DVC, Nextflow, CWL,
+WDL, the R `targets` package, Make, and Calkit all own exactly these concerns:
+a declared dependency graph, an environment attached to each rule, and a
+staleness decision derived from both. ASTRA does not need to reinvent any of
+it. It needs to name the engine and get out of the way.
+
+Underneath both failures is a layering question the spec has never stated
+outright. **ASTRA declares desired state; something else implements it.** A
+command is implementation. A container is implementation. A CPU request is
+implementation. The `container` field is the clearest tell that the line had
+been crossed: the spec started describing *how* to build, without gaining the
+ability to decide *whether* to build.
+
+## Proposal
+
+In plain language: **an ASTRA analysis names the workflow engine that builds
+it, and each output names its rule inside that engine's own workflow
+definition. ASTRA says nothing else about execution.**
+
+Concretely, four changes to `src/astra/schema/analysis.yaml`:
+
+### 1. `Analysis.workflow_engine` — a new enum slot
+
+Names the engine responsible for building this node's outputs. Declared once at
+the root and inherited by every descendant analysis; a sub-analysis may override
+it, which matters most for one extracted to its own directory via `path` and
+carrying its own workflow definition.
+
+The vocabulary is **closed**, listed alphabetically with an explicit note that
+the ordering carries no recommendation:
+
+| Value | Engine | Workflow definition |
+|---|---|---|
+| `calkit` | [Calkit](https://github.com/calkit/calkit) | `calkit.yaml` |
+| `cwl` | [Common Workflow Language](https://www.commonwl.org) | `*.cwl` |
+| `dvc` | [DVC](https://dvc.org) | `dvc.yaml` |
+| `make` | Make | `Makefile` |
+| `nextflow` | [Nextflow](https://www.nextflow.io) | `main.nf` |
+| `snakemake` | [Snakemake](https://snakemake.github.io) | `Snakefile` |
+| `targets` | The R [`targets`](https://books.ropensci.org/targets/) package | `_targets.R` |
+| `wdl` | [Workflow Description Language](https://openwdl.org) | `*.wdl` |
+| `other` | An engine outside this list | name it in the analysis `description` |
+
+The membership criterion is **staleness tracking**: an engine must be able to
+decide, from declared dependencies, whether an existing artifact still reflects
+its inputs. A bare command runner or task launcher (`just`, a shell script, a
+`Makefile` used only as an alias book) does not qualify — admitting those would
+reintroduce exactly the gap this RFC closes.
+
+### 2. `Output.workflow_target` — a new string slot
+
+Names the rule, stage, or target inside the engine's definition that builds this
+output: a Snakemake rule, a DVC or Calkit stage, a Make target, a Nextflow
+process.
+
+Deliberately a bare name. The command, the environment, the resource request,
+and the staleness rule are all read from the workflow definition, where the
+engine can act on them. Several outputs may name the same target; that is the
+normal case when one script emits multiple artifacts.
+
+### 3. Deprecate the execution surface
+
+Not just `command` — the whole implementation surface, as a unit:
+
+| Element | Status |
+|---|---|
+| `Recipe` (class) | Deprecated |
+| `Recipe.command` | Deprecated |
+| `Recipe.container` | Deprecated |
+| `Recipe.resources` | Deprecated |
+| `Resources` (class) | Deprecated |
+| `Analysis.container` | Deprecated |
+
+Deprecating `command` alone would be half a line. A spec that still declares a
+container is a spec that has started implementing, and it would still have no
+answer on staleness. Each element carries a LinkML `deprecated:` metaslot with
+its rationale plus a `DEPRECATED —` lead in its description, so both surface in
+the generated `docs/elements/` reference.
+
+Nothing is removed. Every existing document remains valid.
+
+### 4. Two structural rules on `Output`
+
+- **`from_alias_forbids_workflow_target`** — a re-exported output (`from:`) is a
+  pure pointer; the source already declares how the artifact is built.
+- **`workflow_target_forbids_recipe`** — an output declares how it is built
+  exactly once.
+
+Note that the second forbids the *entire* `recipe`, not merely `recipe.command`.
+This is the layering principle expressed as a constraint: there is no
+half-migrated state where the engine builds the artifact but ASTRA still
+declares its environment.
+
+### 5. One semantic rule for `astra-tools`
+
+`workflow_target` resolves only against an engine, so an output declaring one
+must have a `workflow_engine` in scope at its own node or an ancestor. This
+crosses analysis nodes and cannot be expressed structurally in LinkML; it
+belongs to the semantic validation stage.
+
+## Examples
+
+Before and after, with the environment and resource request moving across the
+line rather than disappearing:
+
+```yaml
+# Before (deprecated)
+outputs:
+  - id: fit_params
+    type: table
+    inputs: [catalog_data]
+    decisions: [fit_method]
+    recipe:
+      command: >-
+        python src/fit_period_luminosity.py
+        --catalog {inputs.catalog_data}
+        --method {decisions.fit_method}
+        --out {output}
+      container: python:3.12-slim
+      resources:
+        cpus: 4
+        memory: "8Gi"
+```
+
+```yaml
+# After
+workflow_engine: snakemake
+
+outputs:
+  - id: fit_params
+    type: table
+    inputs: [catalog_data]
+    decisions: [fit_method]
+    workflow_target: fit_period_luminosity
+```
+
+```python
+# Snakefile — command, environment, and resources live here, where the
+# engine can act on them and decide whether the artifact is stale.
+rule fit_period_luminosity:
+    input: "data/catalog.csv"
+    output: "results/{universe}/fit_params.csv"
+    container: "docker://python:3.12-slim"
+    resources: cpus=4, mem_mb=8192
+    shell: "python src/fit_period_luminosity.py "
+           "--catalog {input} --method {wildcards.fit_method} --out {output}"
+```
+
+`inputs` and `decisions` are untouched. They were always the provenance
+contract — what this artifact depends on, and which choices shape it — and that
+is squarely desired state, on ASTRA's side of the line.
+
+Inheritance and override:
+
+```yaml
+workflow_engine: snakemake        # inherited by every descendant
+
+analyses:
+  validation:
+    workflow_engine: dvc          # this subtree builds with DVC instead
+    outputs:
+      - id: purity_curve
+        type: figure
+        workflow_target: purity_curve
+```
+
+Rejected by the schema:
+
+```yaml
+outputs:
+  - id: result                    # two contradictory answers
+    type: metric
+    workflow_target: evaluate
+    recipe:
+      command: python src/evaluate.py
+
+  - id: reexport                  # a pure pointer cannot also build
+    from: child.metric
+    workflow_target: regenerate
+```
+
+## Implementation implications & migration
+
+**`astra-spec` (this repo).** Implemented on branch `workflow-engine`, pending
+this RFC:
+
+- `src/astra/schema/analysis.yaml`: the `WorkflowEngine` enum, the two new
+  slots, six `deprecated:` annotations, two `Output` rules.
+- `src/astra/datamodel/`: regenerated via `just gen-python`.
+- Docs: `specification.md` gains a "Workflow engines and build targets" section
+  and a migration walkthrough; the Recipe section is reframed as deprecated;
+  `index.md`, `getting-started.md`, `cli.md`, and `README.md` examples
+  converted. `docs/elements/` regenerates via `just gen-doc`.
+- Examples: both `iris` and `iris_pipeline` converted to Snakemake, the latter
+  demonstrating root-level inheritance across sub-analyses.
+- Tests: a valid fixture covering inheritance and override, three invalid
+  fixtures (target-with-recipe, aliased-output-with-target, engine outside the
+  vocabulary). `Analysis-001.yaml` deliberately retains the deprecated recipe
+  path as back-compatibility coverage.
+
+**`astra-tools` (Python CLI + SDK).** Not yet done:
+
+- The semantic check that every `workflow_target` has a `workflow_engine` in
+  scope, with a new error code.
+- `astra init` currently scaffolds a `container:` default and a
+  `recipe.command`; it should emit `workflow_engine` + `workflow_target`, and
+  ideally a starter workflow definition alongside `astra.yaml`.
+- Deprecation warnings on documents still using `recipe`, routed through the
+  existing non-blocking warning channel.
+- `UNDECLARED_TEMPLATE_REF` and the rest of the placeholder-checking logic stay
+  as-is for as long as `recipe.command` survives.
+
+**`astra-typescript` (`@astra-spec/sdk`).** Types and validation mirror the
+Python schema surface and need regenerating for the new enum and slots.
+
+**Compatibility / versioning: minor bump.** The change is additive plus
+deprecation. Nothing is removed, and no document that validates today stops
+validating: the one new prohibition (`workflow_target_forbids_recipe`) can only
+fire on documents using a field that did not previously exist. Deprecation is
+carried as metadata, not enforcement.
+
+**Migration** is mechanical and needs no tooling: move the command into the
+workflow definition, name the rule, replace `recipe:` with `workflow_target:`.
+`container` and `resources` move with it, as does any node-level
+`Analysis.container`. The `{inputs.<id>}` / `{decisions.<id>}` / `{output}`
+placeholders become the engine's own substitution syntax, which every candidate
+engine provides in some form.
+
+**Eventual removal** of `Recipe`, `Resources`, and `Analysis.container` would be
+a major bump, and is explicitly *not* proposed here. This RFC only establishes
+the direction; a later RFC can retire them once real analyses have migrated.
+
+## Questions or objections
+
+Recorded as open. These are the forks this draft expects discussion to resolve.
+
+- **Closed vocabulary, or free string?** A closed enum is machine-checkable and
+  lets the "tracks staleness" criterion actually mean something, but it makes
+  every new engine an RFC (or at least a PR). `other` is the escape hatch,
+  though it is weak: it validates without telling a reader anything, relying on
+  the analysis `description` to carry the real answer. An open string with a
+  documented recommended vocabulary is the obvious alternative. Which failure
+  is worse — an author blocked by the list, or a vocabulary that quietly fills
+  with task runners that don't track staleness?
+
+- **Is forbidding `resources` alongside `workflow_target` too strict?** The
+  strong reading — the one implemented — is that a resource request is
+  implementation and belongs to the engine. But HPC users may reasonably want
+  the request *visible in the spec* as a declaration of what the analysis needs,
+  independent of how any particular engine spells it. Counter-argument: that is
+  a request for a portable resource vocabulary, which is a much larger design
+  problem than this RFC, and `Resources` was never that. Worth deciding
+  explicitly rather than by omission.
+
+  (Note: LinkML forced this to some degree. A rule scoped narrowly to
+  `recipe.command` via a nested `range_expression` postcondition mistranslates
+  in `gen-json-schema` — it emits `'recipe' is a required property`, the
+  opposite of the intent. The whole-`recipe` rule is both correct on the merits
+  and the only form that reliably enforces.)
+
+- **Should `workflow_target` be required on non-aliased outputs?** It is
+  currently optional, so an output may declare neither a target nor a recipe —
+  meaning "no declared build path". That is useful while drafting and for
+  outputs produced by hand, but it is also how an analysis silently stays
+  under-specified, which is the failure mode this RFC is about. Requiring it
+  would be a breaking change and is not proposed; a validator *warning* may be
+  the right middle ground.
+
+- **Where does the artifact land?** `{output}` previously told a reader the path
+  the artifact would be written to. A bare `workflow_target` does not, so
+  answering "where is this figure?" now requires reading the workflow
+  definition. An optional `path` alongside `workflow_target` was considered and
+  rejected as a re-import of implementation — but the loss is real, and matters
+  for `findings` whose evidence points at an artifact.
+
+- **No engine version pinning.** `workflow_engine: snakemake` says nothing about
+  which Snakemake. The workflow definition can pin it, and adding a `version`
+  field would start rebuilding the structured-execution-context object this RFC
+  removes. Recorded so the omission is deliberate rather than overlooked.
+
+- **Self-promotion.** Calkit is authored by this RFC's author. It appears in the
+  vocabulary on the same terms as every other engine — one line, alphabetical
+  ordering, an explicit note that ordering implies no recommendation — and is
+  deliberately *not* used in any example or worked walkthrough; both shipped
+  examples use Snakemake. Flagged here so the choice is on the record and can be
+  challenged.
+
+## References
+
+- [Snakemake](https://snakemake.github.io), [DVC](https://dvc.org),
+  [Nextflow](https://www.nextflow.io), [CWL](https://www.commonwl.org),
+  [WDL](https://openwdl.org), [`targets`](https://books.ropensci.org/targets/),
+  [Calkit](https://github.com/calkit/calkit) — the prior art this RFC delegates
+  to rather than reimplements.
+- [The Turing Way, *Reproducible Environments*](https://book.the-turing-way.org/reproducible-research/renv)
+  — on why recording an invocation without its environment is insufficient.
+- [RFC-0001](0001-establish-the-rfc-process.md) — the process this RFC follows.
+- [RFC-0002](0002-decouple-reports.md) — the same layering instinct applied to
+  reports: ASTRA holds the record and the addressing, an external medium holds
+  the prose. This RFC applies it to execution.
+- Tracking issue: TBD.
diff --git a/rfcs/README.md b/rfcs/README.md
index 173fe457..72ba9459 100644
--- a/rfcs/README.md
+++ b/rfcs/README.md
@@ -54,3 +54,4 @@ discoverability. Each RFC's `status:` frontmatter field is the durable record.
 |---|---|---|
 | [0001](0001-establish-the-rfc-process.md) | Establish the ASTRA RFC process and interim governance | Accepted |
 | [0002](0002-decouple-reports.md) | Decouple analysis reports from astra.yaml | Active |
+| [0003](0003-workflow-engines.md) | Delegate execution to workflow engines; deprecate recipes | Draft |