A governed agentic AI operating system
Your AI writes fast. Can you trust what it wrote?
Fabled turns intent into reviewed, versioned, evidence-backed artifacts — work you can approve, not just accept. Declarative where adaptation creates value; deterministic where guarantees are required.
Release 1 — the Agentic PRD Writer — is in development. The preview shows where it's going.
Terminal transcript simulating the Fabled CLI running the PRD Writer protocol: the Fabled Daemon connects locally, loads the prd-writer protocol with four evaluation contracts, runs an adaptive interview, pauses on a Review Hold for human approval, and exports a versioned PRD artifact package with evidence attached. The transcript is a simulation; Release 1 is in development.
01 — The problem
Agentic AI made output cheap. It made confidence expensive.
Small teams now ship at machine speed — and review at human speed. Something has to give. Usually it's trust.
Untrusted output
Plausible ≠ correct.
The agent's work looks finished. Nobody can say whether it's right. So you re-read everything — or merge on vibes.
Rotting specs
The spec died on day two.
The PRD lives in a doc nobody opens. The code drifted a week ago. Now the source of truth is “ask Priya.”
Evaporating context
Every session starts from zero.
Decisions, constraints, rejected options — gone when the chat closes. Your team's judgment never compounds.
A run log of ungoverned agent work: the agent generated 4,200 lines marked done, chose the architecture unreviewed, assumed the requirements unwritten — and you found out in the pull request, too late.
For a five-person team, this isn't friction. It's the ceiling.
02 — The fix
Better prompts won't fix this. The fix is an operating system: declarative where adaptation creates value, deterministic where guarantees are required.
Better prompts won't fix this. The fix is an operating system: declarative where adaptation creates value, deterministic where guarantees are required.
structured protocols + qualified agents + evaluations + deterministic guardrails
Fabled doesn't promise deterministic AI output. It promises deterministic contracts, governed reasoning, evaluated outcomes — and evidence.
03 — Release 1 · in development
The most expensive bug is the one you specified.
The Agentic PRD Writer interviews you the way a great product partner would — adapting its questions to what you say, pressure-testing the gaps — then drafts a PRD that gets evaluated, reviewed, and approved before a line of code exists. Greenfield-friendly: no repo, no docs required.
- Adaptive interviewit asks the question you didn't think to.
- Evaluated draftthe PRD is scored against contracts, not vibes.
- Human approvalnothing exports until you sign it.
Explore the preview → (mock states are illustrative; Release 1 is in development)
Illustrative product mock: the PRD Writer interview inside a browser window at app.fabled.dev. The user says they are building a usage-based billing service; the interview asks which behavior wins when payment recovery overlaps a plan change, and whether metering is per workspace or per seat; a Review Hold banner shows draft v0.3 awaiting human approval.
The planned flow: Chat, Template, Context, Interview, Draft, then the gates — Evaluate, Review, Approve — and Export.
04 — Governed execution
Approve work. Don't just accept it.
Every Fabled run passes through evaluation contracts and hard gates. A Review Hold stops the line for human judgment. A Policy Block stops it, period. What comes out the other side carries its evidence with it.
Diagram of the planned governance pipeline: intent flows to an agent run, then to evaluations, then to a Review Hold that awaits human approval, then past a Policy Block to an approved artifact. A rejected path leaves the Policy Block with reasons — failure is allowed; unsafe progress is not.
“Failure to merge is acceptable; unsafe merging is not.”
Planned for the Software Factory: six review lenses — context-blind defect, requirements, architecture conformance, security & privacy, deterministic verification, context-aware integration.
Illustration of a planned exported PRD artifact package: a folder containing prd.md at version 0.3 approved, an evaluations folder with four passed contracts, an evidence folder with the attached run log, and an approvals folder with the human sign-off. Beside it, the run's evidence trail: draft v0.1 created, evaluation passed three of four, revision v0.2, Review Hold approved, exported v0.3.
05 — Durable work
Chats evaporate. Artifacts compound.
Fabled's unit of output isn't a message — it's an Artifact Package: versioned, evaluated, approved, with the run's evidence attached. Six months from now, you won't ask what the AI was thinking. You'll open the file.
- Versionedevery revision kept, every diff explainable.
- Evidencedthe reasoning trail ships inside the package.
- Durable runspause on a gate today, resume tomorrow. Nothing resets.
06 — Provider-neutral
Model uplift without workflow rewrites.
Fabled's Agent Harness is a provider-neutral port around the tools you already run — Claude Code and OpenAI Codex, up to launching the real provider TUI inside a managed boundary. When a better model ships, you swap the engine. Your protocols, gates, and artifacts don't move.
- Your toolswraps real provider CLIs; doesn't imitate them.
- Your workflowprotocols outlive any single model.
- Managed boundarythe harness governs what agents can touch.
Harness integrations are roadmap work in progress.
Diagram of the planned provider-neutral Agent Harness: Claude Code and OpenAI Codex orbit the harness as swappable engines, with an empty slot for whatever ships next.
07 — Local-first
Your context is the company. It stays home.
The PRD Writer ships as Fabled Desktop and an interactive CLI on a local Fabled Daemon. Your workspace, artifacts, and evidence live on your machine. Repo context is optional — and read-only.
Architecture diagram: Fabled Desktop and the interactive CLI both talk to a Fabled Daemon that runs on your machine. Only prompts leave for the model provider; your workspace stays home.
08 — The platform
One release of a much larger machine.
The PRD Writer is the first application on FabledOS — the runtime that will govern the full software lifecycle, from intent to release to what your org learns. Same primitives, same gates, all the way down.
The product stack: the Agentic PRD Writer, Release 1 in development, sits on the Fabled Software Factory, the first application, which runs on FabledOS, the runtime and governance layer.
The Software Factory lifecycle: Intent, Requirements — where Release 1 is being built — then Architecture, Plan, Build, Review, Merge, Release, and Learn, all roadmap.
“Fabled is the operating system for that future.”
The primitives underneath
- Protocols
- Artifact Templates
- Artifact Schemas
- Skills
- Evaluation Contracts
- Policies
- Solution Bundles
- Workspaces
- Work Units
- Runs
- Artifacts
- Evidence
- Effects
09 — What's real
No customers. No logos. No launch date theater.
Release 1 is in development. app.fabled.dev is a product preview — you can walk the surfaces, but no live Runs execute yet. We'd rather show you exactly where the line is than blur it.
Policy · no-overclaiming · enforced on this page
- Built & shown
Product preview at app.fabled.dev
This architecture
These gates, as designed
- In development
Agentic PRD Writer (Release 1)
Fabled Desktop + CLI on the local Daemon
- On the roadmap
Agent harness integrations (Claude Code, OpenAI Codex)
Six review lenses
The full Software Factory lifecycle
- Customers
None yet — no logos, no testimonials, on purpose
- Pricing
Being validated with design partners
Everything above this line was labeled accordingly. That's the same standard your artifacts will be held to.
10 — Plans
Priced with design partners, not at them.
Tiers are set. Pricing is being validated with early design partners.
Individual / Founder
For the solo builder who wants specs that survive contact with an agent.
Local-first PRD Writer · your machine, your artifacts.
Explore the preview →Professional / Team
For small product and engineering teams who need shared gates and shared memory.
Everything in Founder · team review and approval flows.
Explore the preview →Enterprise
For organizations that need governance as policy, not convention.
Policies, audit, and evidence at organizational scale.
Explore the preview →Stop accepting. Start approving.
Walk the product preview and see what governed AI work will feel like — gates, evidence, approvals and all.
Release 1 · Agentic PRD Writer · in development