What we learned letting AI agents run whole HDL workflows

Your Agent Says It Is Done

On OCT 14, we handed AI agents two complete engineering objectives and let them work. Both finished. In the first workflow, deterministic analysis still had findings after the agent called the job done.

From one task to the whole job

Most experience with coding agents so far is task-sized: write this module, fix this include, format this file. You prompt, you check, you prompt again.

Agentic engineering changes the unit of work. The agent gets an objective, breaks it into steps, chooses which tools to call, and carries context from one step to the next until it decides the work is finished. Nobody reviews between the steps.

That raises one practical question: “what is the agent’s done worth?” Our Agentic Engineering webinar tested it on two workflows, live.

Workflow 1: from datasheets to a complete design

The agent, Claude Code in the demo, started from a folder with nothing but requirements and datasheets for an image acquisition system: capture a frame from an imaging detector, store it in DDR2 memory, and send it to a host over a USB FIFO interface on command.

It loaded a set of Sigasi skills and built the design, calling the CLI throughout. The CLI ran in daemon mode, which keeps the project loaded between calls, so each check only processes what changed.

What Sigasi caught while the agent worked

During the build, the agent introduced, and then fixed from Sigasi’s diagnostics:

• declaration problems
• parse errors
• variables used as the target of a signal assignment
• an incomplete reset branch

None of these are style issues. They are the kind of mistake that looks plausible in text and fails in the project.

What was still there after “done”

When we asked the agent to analyze the finished design, Sigasi still reported unused signals and variables, problems in generic mappings, unconstrained signal types, and null ranges. None of them stopped the agent, so it classed them as non-critical and left them in.

That is the lesson of the first demo. The agent’s “done” means the agent stopped. Deterministic analysis shows what is still worth an engineer’s attention.

Workflow 2: from an existing codebase to a simulation

The second workflow started from the open-source CVA6 repository. The agent read the documentation, identified the dependencies, and configured a Modular Project. It then asked Sigasi for the compile order and a compilation recipe, which adds library mappings, language levels and preprocessor definitions, and used that recipe to write the scripts to compile and simulate the design with Questa.

The point is where the setup came from. Without the recipe, the agent would reconstruct the compilation setup from source every time a script needs to change. With it, the setup comes from the project, deterministically, and the agent reuses it. Processing less to get there can also reduce token usage.

Skills tell the agent how. The CLI tells it what is true.

Both workflows rested on the same split:

• Skills are reusable instructions for a kind of work: how to set up a project, when to check diagnostics, how to handle compile order. They save explaining the process again in every prompt.
• The CLI gives the agent the same project-aware analysis Sigasi Visual HDL gives engineers in the editor, from the command line, as one tool among several in a scripted workflow.

A skill does not execute anything. The agent does, and it calls the CLI when it needs a fact about the project.

A policy that works when agents run whole workflows

If you are deciding how far to let agents go, this is the setup the webinar supports:

• Let the agent work the whole objective, not one prompt at a time.
• Give it skills for the process and the CLI for project facts.
• Check while it works, and again after it says it is done.
• Let the engineer decide what to do with the remaining findings, and write recurring decisions into a skill so the agent handles them next time.

What this does not claim

The agent wrote the HDL. Sigasi analyzed it when the agent called it. A clean list of diagnostics means the checks that ran found nothing over the scope they cover. It does not mean the image acquisition design works: that still takes simulation and verification, and the engineer stays accountable for what moves forward.

Two workflows on two projects are a demonstration, not a benchmark. Your project, agent and skills will produce different results.

Watch the webinar on demand

See both workflows end to end, including the agent’s own reports and the findings left after it declared the job done. Fill in the form to get instant access to the recording.

We do not recommend using LLM agents for RTL without a deterministic check by Sigasi that keeps pace with the agent.

AI can generate HDL fast. Sigasi keeps engineers in control.

2026-10-15