Skip to content

How Daedyl works

A short guide to working in Daedyl: projects and stages, running engineering tools, Daedyln and Autopilot, and how every result keeps its history.

Getting started

Sign in once from the homepage. Daedyl opens your last active project at Build, or your projects list when you have none yet.

A project holds one piece of engineering work: its requirements, designs, runs, results, decisions and approvals. Create one from Projects with a name and a preset, for example an RF amplifier or a digital chip.

Each project's Overview shows where it stands and a Next line: the one thing to do now, with a button to it and a question you can hand straight to Daedyln.

The six stages

  1. Define. Answer six guided questions (what you are defining, the outcome, who needs it, constraints, what counts as evidence, what must be ready), or describe the project to Daedyln and accept the context it proposes. Daedyl turns that into requirements and architecture drafts you read, edit and accept.
  2. Design. Schematics, architecture diagrams and constraints, with Daedyln beside the object you select.
  3. Build. RTL checks and simulation, synthesis, place and route, timing, layout, DRC and LVS, formal proofs and equivalence, through governed skills and workflows.
  4. Simulate. Circuit and RF analysis (S-parameters, noise, stability), electromagnetic, thermal and device analysis.
  5. Optimize. Sweeps and design-space exploration: compare candidates on the numbers that matter and propose one back to the design.
  6. Evidence & approvals. Every record, its provenance and the approvals on it; the definition package is generated from an approved definition and never changes afterwards.

Skills, runs and workflows

Engineering tools run through skills: a reviewed, versioned permission for a project to use a tool. A project can only run tools it has an active skill for; you set them up on Build > Skills.

A run is one tool on your inputs. A workflow is a chain of runs where each step reads the previous one's result, for example RTL to GDS, device to compact model, board design, a fault campaign or a cell library. Each run records its inputs, the engine and its version, its log and its result.

Daedyln

Daedyln is the AI engineering agent inside every project. Open it from any screen (Ctrl+J). It answers from the project's records, with the numbers as recorded, and says when the records do not say.

Ask it to run something and it proposes the exact run or workflow and inputs; nothing starts until you press Start. When a run finishes, it explains the result; when one fails, it explains why and what to change.

By default Daedyln uses a model running on Daedyl's own hardware, so nothing leaves it. A founder or administrator can add a cloud model key (Anthropic, OpenAI or OpenRouter) in Settings, and a project's founder or administrator can choose it for that project, confirming that its records go to that provider.

Autonomy and Autopilot

Each project has an autonomy level, set by its founders and administrators on the Autopilot page:

  • Assist (the default): Daedyln answers and proposes; you do every step.
  • Co-pilot: the Next step gets a "Let Daedyln do it" button that takes that one step.
  • Autopilot: give it a goal and Daedyln works the project forward within a budget of runs, time and cloud cost. It answers the Define questions (marked as inferred, never as fact; your own answers are never overwritten), prepares the drafts, and chooses and runs what the project can run, reading each result before the next. It stops for you at acceptance and approval, and ends with a summary.
  • Full autonomy: as Autopilot, and Daedyln may also accept drafts and approve the definition in that one project, under a delegation a founder grants. Every such approval says it is the AI's. Lowering the level ends it.

You also choose how deep it goes, how far from known designs it may go, and what evidence it needs before moving on.

The council. Autopilot can put one to four models on each decision. Each answers alone, then reads the others' answers without knowing which model wrote them, names the strongest idea and the weakest assumption, and revises. A chair writes one agreed plan, and anything they still disagree on becomes a run to settle it. Every message is kept on the run for you to read.

Evidence and approvals

Every result says what kind of evidence it is: estimated, simulated or measured. A simulated number is never shown as a measured one.

Results link back to what produced them: requirements, design decisions, the run and its inputs, the engine and version, and forward to the approvals that rest on them. The evidence graph lets you trace any requirement down to its results and see where a claim is only assumed.

Approvals need a person with the right role, recorded with who, when and exactly which evidence they approved. If that evidence changes, the approval no longer holds and a new one is needed.

Security and your data

Daedyl runs on hardware its team owns, on a private network. Sign-in is single sign-on; every request is checked against the project and the person's role, and policy decides what each role may run.

Changes, runs, approvals and Daedyln's own actions go into an append-only audit history. Keys for cloud models are kept in a secrets vault and never shown again after they are saved.