Agent architecture and responsibilities

Five jobs, each done by its own agent, and the one rule they all share.

StudentEngineerDeveloper 6 min read · Updated 2026-10-31

        On this page
        The five jobsThe rule they shareWhat it cannot doWhat it costsRead onDesign Chat and the V&V evidenceHow the Agent keeps learningWhere the Agent's references come fromThe "write RTL this lab builds cleanly" skill packWhere the Agent can be asked
      

Nelix Agent is the lab's AI engineer. It is not one chatbot. It is a set of agents, each with a narrow job.

The five jobs#

Agent	What it does	Where you meet it
Design Chat	Turns what you ask for into a specification, a design and a test. Answers questions about your project	The Design Chat page
Verification	Writes a test for your design	The Testbench tab
Stage explainer	Reads a failed run and says what failed, the evidence and one next step	Ask Nelix Agent on a failed stage card
Timing and implementation	Explains measured build results and recommends fixes	A build that did not meet timing
Diagnosis and repair	Finds the cause of a failed simulation and proposes the smallest change that fixes it	Debug with Nelix Agent on a failed card, or `nelix debug`

Nelix Agent may only state what the tools measured. A passing test, a cell count, a timing margin: it reports these when the run's results say so, and names the run. It cannot invent a number.

The parts of the lab. The language model answers the AI engineer; the platform runs the design tools and shows it the results. In Design Chat it can also read the open project's files, runs and reports, and search the web; the platform runs those lookups for it.

What it cannot do#

It cannot change your files. It proposes; you push, or accept and merge.
It cannot run commands from inside your project.
In Design Chat only, it can search the public web (for example for a datasheet or a protocol detail). The search runs on the AI provider's servers, and the search words can come from what you wrote, so don't put anything confidential in a Design Chat request.
It cannot see another person's project.
It cannot weaken a test, change an expected value or relax a timing target to make a run pass.

What it costs#

Design Chat messages, Ask Nelix Agent, Debug with Nelix Agent and generated tests use AI credits from your plan; a longer answer uses more. Reading a failed card does not. See Plans and allowances.

Read on#

Design Chat and the V&V evidence#

Design Chat has two read-only tools for a project's verification and validation: vv_checks (the V&V tab as data: every check with its measured state, run id, details and report links) and vv_file (one kept report, log or export). The rules it follows: measured states are authoritative and never promised or changed by the Agent; a failing check is explained from its evidence (the detail line, the actual value, the signal or cell named) before any recommendation; a hard tapeout gate is never "waived" by advice; a "not run" row is explained, not guessed. Every V&V page carries an Ask Nelix Agent button that opens Design Chat with the question for that row. A third read-only tool, teaching_context, gives the Agent the person's own teaching side: their roles (university admin, teacher, student) and classes, the lab's grading rules (the mark parts and weights, the visible-test gate, late work, the hint ladder and its penalties: a mark comes from the grader's measured results only), the AI levels and where they apply, and their assignments (state, due dates, counts to review, or, for a student, their own status and what the Agent may do while the assignment is open); it never shows another person's data or a hidden test. A fourth read-only tool, knowledge_lookup, searches the lab's knowledge library: short notes on what each V&V check measures and does not cover, tool error texts at the lab's versions, what each process runs or cannot run and why, the Agent's own rules, lessons the lab learned, and verified fix records; the Agent cites a note as [K1], and the project's measured results always come first.

How the Agent keeps learning#

The Agent's reference notes come from a knowledge library that grows from the lab's own work, in four ways: repairs (a fix the AI engine verified on a copy becomes a fix record: the error kind, its cause and what fixed it); V&V outcomes (a check that failed and later passed on the same project becomes a lesson: what was seen, why, and what to check first; a check that stays failing for more than a day over several reruns is listed as a sticking point); Design Chat (an answer that read the V&V evidence and named a check which then passed within a day becomes a lesson, written from the run's own texts, never from what the person wrote); and the lab's own notes (lessons the owner writes). Every record is written in general words only: no module, signal, file or project name, no e-mail, path or id, and a deterministic check refuses a record that contains one. A new record stays pending until the owner approves it; only approved records are offered to everyone. Anyone can be kept out of the library (an opt-out list), in which case records made from their work are removed. Drafting these records is paid by the lab, never from a student's AI credits, and no note ever outranks a measured result.

Where the Agent's references come from#

The reference notes the Agent may cite ([K1], [K2]...) come from a curated library, never from a live web search. Every source in it was named by the lab's team, fetched once at a pinned version (a tool's tag or commit, or the date a page was read), and recorded with its licence and how it may be used: full text for permissive licences, short credited excerpts for copyleft documentation, and titles with links only where a licence does not allow storing the text. Sources marked for it are fetched again once a week (Sunday, 04:00 UTC) and only the pages whose text changed are replaced. The Agent itself has no web access when it reads these notes: it sees only what the library holds, at the version recorded, and a note never outranks the measured results of your run. The sources used in an answer are listed under it with their licence.

The "write RTL this lab builds cleanly" skill pack#

Since 31 October 2026 the Agent carries a skill pack of this lab's measured facts about what builds cleanly: one clock and where its period lives, reset conventions, the antenna repair levels, size and time expectations for starters and for a 7,000-cell design, and what every V&V row's words mean. Every fact in it was measured on the lab's own server (the PicoRV32 benchmark and the starters), with the date. The pack's layout, one skill with short parts read when the task needs them, follows the Claude Code skills that LiveHD (UC Santa Cruz MASC group, BSD 3-Clause) ships for its own hardware language; LiveHD's authors do not endorse this lab.

Where the Agent can be asked#

Besides the Design Chat page itself, an Ask Nelix Agent button opens Design Chat with a question already written about what that page shows: every V&V row and the power-on page; a project's Requirements page (specification and constraints); the Overview's next step (architecture guidance, when the page is not already offering a debug); Results and Compare builds (power, performance and area); the University Overview, Classes and Teachers pages (roles); the teacher's assignment list, marks, edit and student review pages (deterministic grading); the student's assignment page (with one line on what the Agent will and will not do while the assignment is open); and, in the Process Lab, a course module and the yield analytics part. Pages without a project of their own ask about the person's most recent project. The Agent explains and recommends; measured results and marks are never changed by it.

Something unclear or missing? Send feedbackSign in and use Feedback to tell us.