Where agents sit in chip design today, an agent architecture for a design-services house, and a working agent tested on real OpenROAD runs and real user questions.
A Penang IC design-services house, Bursa-listed, about 350 engineers, moving from billable design hours toward turnkey AI-chip work and its own IP.
Each engineer-hour saved on run triage, script writing and spec lookup goes straight to margin. Customer design data stays inside the building, so models run on-prem.
An agent that reads the OpenROAD docs, parses real timing reports, proposes a repair, runs it in OpenROAD and accepts it only if timing and area checks pass.
| Takeaway | Why it matters |
|---|---|
| Test on real questions | My own 32 questions gave 91% hit@1. 34 real questions from OpenROAD users gave 35%. Golden sets must come from the engineers who will use the agent. |
| Execute, then believe | Every model step ran in OpenROAD and was re-measured. A step that ran but made TNS worse was rejected, as was a hold fix applied where hold was not violated. |
| Guardrails beat model size | On the hold case a fixed rule and a local 7B model closed hold; the hosted 70B model did not. The verifier blocked 7 of 24 proposed steps before they ran. |
| Local models are enough to start | Qwen2.5 7B on a laptop CPU closed the hold case with no data leaving the machine. That is the shape of an NDA-safe deployment. |
| Know when to escalate | At a 0.25 ns clock no repair move fixed setup. The agent said so and handed the call to a PD engineer. |
Public facts only, each checkable at the link.
| Item | Public fact | Source |
|---|---|---|
| Company | Oppstar Berhad, Bursa ACE Market (0275), listed December 2022. Founded 2014 by former Intel Malaysia engineers. HQ Bayan Lepas, Penang; office in Kuala Lumpur. | oppstar.com.my |
| Group | Oppstar Technology, Oppstar Microelectronics, AIRIS Labs (AI unit), Oppstar (Shanghai), Oppstar Japan; Taiwan office from January 2026. | contact page, The Edge |
| People | About 350 staff, up from about 180 five years ago. | The Edge, Sep 2026 |
| Direction | "Oppstar 2.0": turnkey design, own IP and edge-AI SoC work. Arm Flexible Access token for a 12 nm edge-AI SoC; Rapidus CORE Design Solution Associate (Oct 2026). | The Edge, The Star |
| AI chip work | Reconfigurable AI chip contract for a Japanese client; MoUs with Inventec and its NPU unit AsicAI (Sep 2025). | Digitimes, NST |
| Results | 1QFY27 (to June 2026): revenue RM15.9m, net profit RM4.1m, after seven loss quarters. Net cash, no debt. | DNA |
| Hiring now | Eight open roles; this is the only software or AI seat. The others are physical design, SI/PI, firmware, validation and warehouse. | directory card |
Where agents sit in chip design in 2026. Malaysian peers from my Malaysia semiconductor directory (133 companies with a Malaysian site).
| Player | What they ship | Gap for a services house |
|---|---|---|
| Synopsys | AgentEngineer, a multi-agent workflow on Synopsys.ai Copilot and DSO.ai | Tied to one vendor flow and its licences; customer projects mix vendors |
| Cadence | ChipStack AI Super Agent (Feb 2026), verification first; Cerebrus and JedAI | Same: per-vendor, per-seat |
| Siemens | Fuse EDA AI Agent: a supervisor that calls tools and checks itself with deterministic tools | Closest to the verify-by-tool pattern used here; Calibre-centred |
| Cognichip, ChipAgents, Ricursive | Venture-backed AI-for-chip-design startups (2026 rounds of $50m to $300m) | Product companies selling seats; customer data leaves for their cloud or their models |
| Malaysian design houses | SkyeChip, Infinecs, GreatAsic, Alphaswift: peers in Penang and Kuala Lumpur, several also Rapidus CORE associates | Same talent pool and the same margin pressure; none shows an in-house agent practice publicly |
| Oppstar opening | Vendor-neutral agents around the flow: log and report triage, script drafting with verification, spec and PDK lookup, run on-prem per customer | Uses the vendor tools as tools, keeps customer data inside, measured on its own engineers' tasks |
Numbers from the demo's Evals and Real OpenROAD runs tabs; traces with input hashes are downloadable there. Open the demo ↗
| Finding | Evidence | Implication |
|---|---|---|
| Self-written evals flatter | Dense retrieval: 91% hit@1 on 32 questions I wrote; 35% on 34 real questions from OpenROAD GitHub Discussions, labelled by the commands in maintainers' replies and frozen before tuning | Build the golden set from the team's own chat and tickets in week one |
| No retriever wins on 34 questions | Hybrid vs dense vs BM25: no difference at p < 0.05 (exact McNemar). Registry routing plus hybrid had the best hit@3, 62% | Pick on cost and citations; collect more real questions before tuning |
| Setup at 0.25 ns is not repairable by sizing | gcd on Nangate45 at a 0.25 ns clock: setup WNS -0.154 ns, TNS -6.84 ns. Every repair_timing variant from 3 backends moved TNS by under 1% | The agent escalates: check report_clock_min_period (0.41 ns here) or restructure |
| Hold is repairable, with care | With a 0.1 ns hold margin, plain repair_timing -hold closed hold (+0.007 ns) at 1.5% area. The 70B model asked for a 0.1 ns hold margin and broke the area budget | Assertions on area and the other timing class catch over-fixing |
| The verifier earns its place | 7 of 24 proposed steps (6 of 16 from the two LLMs) were blocked before running: setup and hold mixed in one step, a fix for a class with no violation, malformed Tcl, repeats | Never run model Tcl unchecked |
| Long context is not free | Top-5 chunks: about 650 tokens per question. Full OpenROAD docs: about 120K, which does not fit a 32K on-prem model | Retrieval stays, as a tool the agent calls |
| JD duty | How I would do it | Detailed in |
|---|---|---|
| Design multi-step reasoning flows with tools and APIs | Plan, call tools, answer, verify, one retry; plans checked against each tool's input schema | demo, §04 |
| Prompt and policy optimisation; evals for success, safety, reliability | Numbered policy rules; evals per change on a frozen set of real questions; safety as blocked-step and false-alarm rates | ★, §04 |
| LLM platforms, RAG and orchestration | Open-weight model on-prem (Ollama or vLLM); hybrid retrieval with heading-aware chunks; custom loop, LangChain or LlamaIndex where they save time | §04 |
| Translate business and engineering use cases into agents | Start from the task an engineer repeats daily; measure its time before building | §05 |
| Encode workflows with domain experts | PD leads label 50 real questions and 10 past runs with known fixes; their rules become allowlists and assertions | §05 |
| Connect agents to EDA tools, databases, CRM, microservices | Tools run in batch on a checkpoint copy in a sandbox, with timeouts and no network; outputs parsed into JSON | demo |
| Multi-agent coordination and workflow automation | Separate triage, repair and report agents only when one loop is measured to be the bottleneck | §04 |
| Evaluation: task success, guardrails, reliability, feedback | Task success from re-measured tool output; guardrail eval with planted errors; thumbs-down feeds the golden set; A/B on engineer time | ★ |
The same loop as the demo, sized for real projects.
| Layer | Design | Why |
|---|---|---|
| Data boundary | Index per customer project; access checked at retrieval; models run on-prem; every retrieved chunk logged | Customer RTL, netlists and PDKs are under NDA |
| Knowledge | Tool manuals, PDK docs, team wiki, past run logs. Heading-aware chunks with the heading path kept; BM25 plus embeddings; command registry from the tools' own definitions | Engineers ask by command name and by symptom |
| Tools | Report parsers (timing, DRC, LVS, regression logs), registry lookup, batch tool runs on checkpoint copies | Numbers come from tools; the model never recomputes slack |
| Verifier | Allowlist per checkpoint stage; flags and values checked against the registry; Tcl syntax limits; repeat and wrong-target checks | A real but wrong command costs a grid slot and an engineer's trust |
| Execution and assertions | Run, re-measure, accept only if the target metric improves and other metrics stay inside budget | Ran is not the same as fixed |
| Evals | Frozen real-question set, past-run replay with known fixes, planted-error guardrail set; results per model and per prompt version | Model and prompt changes ship on evidence |
| Trace and feedback | Hashes of inputs, tool and model versions, every step and check, exported as JSON; engineer thumbs-down adds a case | Audit for customers; the eval set grows from use |
Sit with two PD and verification engineers. List the ten tasks they repeat most and time them. Agree the data boundary with IT. Stand up an on-prem model and a first golden set of 50 real questions.
Ship one agent for the top task, likely run and report triage, read-only. Weekly evals on the golden set. Engineers rate every answer.
Add verified script drafting with sandboxed dry runs. Report hours saved against the baseline. Pick the second workflow, possibly the CRM side, by measured demand.
Role: AI Agent Engineer, Oppstar Technology, Penang (public job posting, 2026).
Company: oppstar.com.my · The Edge, "Oppstar 2.0" · Digital News Asia · The Star · Malaysia semiconductor directory.
Agentic EDA: Tom's Hardware, 2026 · The Register on ChipStack · TechCrunch on Cognichip.
Work sample: OpenROAD module docs (BSD-3-Clause) and OpenSTA command definitions; design gcd on Nangate45 from OpenROAD-flow-scripts (Docker image openroad/orfs); real questions from OpenROAD GitHub Discussions. Models: Qwen2.5 7B via Ollama (local), Llama 3.3 70B on Cloudflare Workers AI. The demo and the Python CLI are my own work, built for this application.
Method: the retrieval and agent comparison is also published as a reusable calculator in the Agent Harness Lab, Retrieval tab.
Demo: oppstar-agent.pages.dev
Independent homework for the Oppstar Technology AI Agent Engineer role · 2026 · edwardtay.com