Describe the process. Read the manifest.
The workflows that matter most are the ones too conditional to script and too important to hand to an opaque agent. Miriel compiles a plain-language description into a typed, reviewable application manifest — with explicit grants, simulation before activation, and a rollback.
// reviewable before it runs, versioned after
The choice today is a brittle script or an agent nobody can review.
Real back-office processes are conditional, span several systems, and need a human at two or three specific points. Scripting them means encoding every branch by hand and maintaining it forever. Handing them to a general-purpose agent means a loop of tool calls nobody can read before it runs, nobody can diff after it changes, and nobody can roll back when it goes wrong at 2am.
The missing piece is not more autonomy. It is a reviewable artifact — something a person can read, reason about, test, version and revoke, that still came out of a sentence rather than a sprint.
So the output of the natural language is a manifest, not a transcript.
- Typed steps, not free-form tool calls. A workflow is built from a fixed vocabulary — query, classify or extract, reconcile, request input, await event, call tool, emit event — so the shape of what will happen is legible before anything happens.
- Triggers that include your data. Manual, cron, or on new ingested data — so a process can start when the document lands rather than when a poller next wakes up.
- Explicit grants. The manifest names exactly what the app may touch. Nothing is discovered at runtime, and the knowledge graph inherits the same boundaries rather than creating new exposure: cross-document insight only emerges from material the asker was already entitled to see.
- Simulate before activation. Run it against real data with the side effects held back, read what it would have done, then turn it on.
- Amend in natural language, version everything. Change it by describing the change; the diff is against the manifest, and a bad version rolls back whole.
- Answers that survive the handoff. Because every step reads from the same store, a reconciliation step can join a PDF clause, a warehouse event and a call transcript in one query — and every output carries its sources, so the human at the "request input" step sees why they were asked.
- “Every morning, reconcile yesterday's open claims against the policy documents. Anything that does not match, ask Dana.”
- “When a new inspection report lands, check it against the governing spec and flag any parameter out of range.”
- “Watch for purchase orders that arrived without a matching receipt for more than 10 days, and open a ticket.”
- “Summarize what changed across these three drawings since revision B, and who signed each change off.”
Every answer comes back with its evidence — the clip, the row, the clause — so it can be checked rather than trusted.
Code-size comparison is against hand-built retrieval and orchestration pipelines doing equivalent work. The point is maintainability: code any engineer on the team can read, instead of a brittle mess of generated glue.
Compilation is good at processes that are genuinely describable. A workflow whose real rules live only in one person's head will compile into a manifest that is confidently wrong, and the fix is writing the rules down — which is work, and worth doing regardless. We would scope a first app around a process that already has a written procedure.
GET STARTED
Bring us a question you cannot answer today.
The fastest way to find out whether this fits is to pick one real question your current stack cannot answer, and see what it takes.
// no spam, just a beta waitlist