Skip to main content
GatiFlowIntelligence Platform
Deep DiveLoginRegister

GatiFlow Intelligence Platform

TermsPrivacyComplianceOpt-outMethodologyAcademyChangelogStatus

← All Deep Dives
Saturday Deep Dive

The RTL Agent Stack Is Converging Before the Mention-Count Data Notices

Saturday edition · November 14, 2026 · 2227 words · 12 min read

Since April, at least three independent groups have published agentic systems that put a language model inside a loop of real chip-design tools: Dr. RTL from the Hong Kong University of Science and Technology in April, RTLScout from Huawei in June, and ChipOmni on November 9 from a twelve-author team spanning the University of Science and Technology of China, Stanford, City University of Hong Kong, Tongji University and TeraPrime. On November 10 the Dr. RTL group added an invited ASP-DAC 2027 overview that sorts its own program into three connected directions: realistic benchmarking, tool-grounded RTL optimization and design-specific synthesis optimization. That is three independent teams and a fourth document from the first of them. According to GatiFlow intelligence, the overview entered tracking on November 11 with 5 mentions from our arXiv collector, logged as newly active and single-source, and ChipOmni arrived the same day with 4. A mention-count dashboard reads that profile as noise.

In the same 24 hours, the only repository in GatiFlow's GitHub velocity ranking was deepseek-ai/dsh-superagent-prototype, a research prototype of a proactive coding agent that DeepSeek's Harness team open-sourced on October 31 and that Business Standard framed as a challenge to Anthropic's Claude Code. Our GitHub collector had tracked it for 11 days by the November 11 close, when it stood at 22,484 stars, 10.3 percent above its average over the preceding week. That figure is a gap against an average, not a growth rate, and the growth rate is falling: according to GatiFlow intelligence, the repository added 20 stars in its latest six-hour reading, a 0.1 percent move, and the final daily synthesis of the cycle flagged the first sign of deceleration while cautioning against using the repository as a proxy for the health of the space. In that same synthesis, the two EDA papers sat among five single-paper arXiv entries that it judged did not cluster into a second theme. A general-purpose coding agent is winning the attention metric, and a convergent pattern in domain-specific hardware agents is not.

Why does this matter on a cycle of weeks rather than a single news day? Because the pattern is only legible when the papers are stacked across months, and the stack is wider than three groups. The overview cites multi-agent and evolutionary RTL optimizers from other labs, among them CODMAS, POET and EvolVE. Read together, these systems share a recipe: ground the agent in real EDA tools instead of asking it to write hardware from a blank prompt, gate every candidate with a correctness check, and keep a memory of what worked.

Dr. RTL is the most explicit about it. According to its paper, a multi-agent loop built on the Claude Code command-line interface, with Claude Opus as the backbone, works against Synopsys Design Compiler, only candidates that pass a Cadence JasperGold sequential equivalence check can be promoted, and a library of 47 pattern-strategy entries distilled from earlier runs guides new designs. Across 20 open-source designs averaging 812 lines, it reports average worst and total negative slack improvements of 21.3 and 16.9 percent with a 5.8 percent area reduction over commercial synthesis alone, for roughly a week of wall-clock time and about $50 of model usage. Its ablations show where the gain comes from: dropping register-level slack feedback cut the improvement to 9 and 9 percent, collapsing to a single agent cut it to 10 and 7, and dropping the skill library cut it to 14 and 11, so carrying experience forward is the smallest of the three levers.

RTLScout reaches the same shape by other means: its agent rewrites designs in a Python-embedded hardware language against Yosys and OpenROAD feedback, a self-checking simulation gates each result, and an elite pool of past designs plus written lessons seeds later runs. On a 16-bit floating-point multiplier it cut area by 34 percent and delay by 38 percent, and its Pareto front beat a published multiplier synthesized with Cadence Genus on the same process. ChipOmni, whose paper mentions neither Dr. RTL nor RTLScout, applies a verification-driven loop with a debugging memory to generation instead of optimization, across analog circuits, RTL, testbenches and formal assertions, and reports solving all 30 AnalogCoder-Pro tasks and, in one long-horizon case study, taking a RISC-V specification to a micro-SoC layout that passes physical sign-off. The tasks differ; the architecture does not.

The decision space for anyone building here splits into three alternatives, and conflating them is the most common mistake. The first is a general-purpose coding agent pointed at hardware description with little scaffolding: cheap and flexible, and weakest where it counts. In the overview's RTL-OPT benchmark of 36 tasks, single-shot models such as GPT-4o-mini, Gemini 2.5, DeepSeek V3 and DeepSeek R1 beat the unoptimized input on at most 41.7 percent of tasks and the human-optimized reference on at most 13.9 percent. On original human-written RTL under commercial synthesis, a single-shot Claude Opus improved worst negative slack by 2.4 percent, against 21.3 percent for Dr. RTL's loop. A July benchmark that ran agents through a commercial RTL-to-GDS flow on the PicoRV32 core found a related weakness among agents that do use tools: domain-specific EDA skills improved their grasp of individual subtasks without ensuring that the long-horizon flow completed, agents with similar design progress differed by up to 141 times in cost efficiency, and tool-interface mismatches, notably Tcl commands that depend on the tool version or execution mode, were a major source of physical design failures.

The second alternative is the vendor platform, which wraps the same loop around proprietary tools. Cadence calls its ChipStack AI Super Agent, in early access since February, the world's first agentic workflow for chip design and verification, lists Altera, NVIDIA, Qualcomm and Tenstorrent as early deployers, and quotes Tenstorrent cutting verification time by up to 4X on three design blocks in a three-month evaluation. Synopsys says more than 50 customer engagements are underway for AgentEngineer and its Autopilot Platform, with general availability planned for the end of 2026, and describes the platform as pairing ground-truth EDA engines that validate outcomes with reusable skills and persistent memory, which is the research recipe in vendor packaging. In July it put two autonomous workflows on Microsoft Discovery, where AMD is evaluating them and Synopsys reports early debug-cycle reductions of 25 to 40 percent. Siemens launched its Fuse EDA AI Agent in March. Every figure in this lane is vendor- or customer-reported.

The third alternative is the published research lane, and it is not uniformly open. Dr. RTL's loop runs on Claude Opus with Synopsys and Cadence tools as its oracles, and it publishes its dataset, evaluation environment, agent code and learned skills on GitHub; RTLScout's loop runs entirely on open-source tools, Yosys, OpenROAD and Verilator among them, plus an open-weights model, and its code is public too. Both state their baselines and methods, so a buyer can check the numbers. The cost is scale: Dr. RTL's 20 designs are modest next to a production SoC, and RTLScout's results come from four IP blocks and 14 small benchmark cases.

Read across the lanes, the sources point in different directions. GitHub, in GatiFlow's velocity data, shows one general-purpose coding agent and no EDA agent repository among those our collector tracked. arXiv shows three independent groups and a fourth document converging on tool-grounded, verification-gated loops, and the two November papers sit in our collector as newly active, single-source entries. Commercial sources show Cadence, Synopsys and Siemens productizing agent workflows built on their own tools, while ChipAgents, which said in February that it was deployed at 80 semiconductor companies, added a $60 million extension led by B Capital that Reuters reported on July 29, bringing its Series A total to $131 million. Vocabulary may be part of why counters miss the pattern: Cadence sells a super agent, Synopsys an AgentEngineer, ChipAgents an agentic workforce, and researchers say tool-grounded or self-improving agents, so a tracker keyed to one label sees a fragment.

Hiring is the quiet layer. Across the past month of reports, none of the postings our hiring collectors captured names RTL, Verilog, VHDL, SystemVerilog, EDA, ASIC, FPGA or VLSI in its title, role, stack or description excerpt, while postings with agent or agentic in the title appear throughout. Chip companies are present: NVIDIA, Qualcomm and Altera Semiconductor post for generic AI-platform and agentic-AI roles, and they are three of the four companies Cadence names as early deployers. The capability has not yet surfaced as its own job category in the horizontal boards we cover.

The contrarian read is that the market is over-rotating on vendor autonomy labels and under-rotating on a less marketable fact: the same architecture is now documented in public papers with stated baselines and, in RTLScout's case, an entirely open toolchain. The labels do not line up. Cadence told Computex its agent reached Level 5, and its own strategy director described the demonstration as Level 5 within a single bounded domain, with most of the technology closer to Level 4. Synopsys classifies AgentEngineer as Level 5 as well, on a scale that keeps engineers in the loop, and a September survey of the three vendors reads such ladders as marketing scaffolding around one underlying loop.

A skeptic can find reasons to look away from the research, too. An automated arXiv triage run by the pact-safety touchstone project, which tracks standards and benchmarks for general-purpose models and agents, set the HKUST overview aside as a narrow hardware domain outside general-purpose LLM and VLM capabilities. The filter did its job, and it shows how a general-purpose lens, like a mention-count dashboard, pushes domain work to the margin. Domain work also has something general-purpose work lacks: a hard oracle. Equivalence checking rejects rewrites that change behavior, and Dr. RTL reports an average sequential-equivalence pass rate of 86 percent, so the loop's failures are visible rather than silent. That makes RTL agents easier to audit than most agentic coding work, though 20 designs and 14 small cases show a direction, not production readiness.

If you are building in or around EDA tooling, chip-design infrastructure or verification tooling for safety-critical software, this is the conversation to have with your lead or portfolio company in the next two to four weeks: decide which of three bets you are making. A vendor platform from Cadence, Synopsys or Siemens buys integration and inherits vendor timelines. The open tool-grounded lane means building the loop yourself on open or licensed tools and benchmarking it against a commercial baseline. An independent funded layer such as ChipAgents sells across vendors and has to earn trust inside programs where the incumbents bundle their own agents. What separates the three is who owns the verification oracle, so ask any team in this space which check gates each candidate, who licenses it, and what survives on production-scale blocks. Then ask how much of the result comes from the harness around the model rather than from the model or its memory; the July RTL-to-GDS benchmark concludes that robust agentic EDA needs controlled execution and process-level evaluation along with structured tool interfaces and persistent design context, not only stronger models.

Vendor commitments cluster at year-end. Synopsys has said general availability for AgentEngineer and the Autopilot Platform is planned for the end of 2026. Cadence has promised customer early access to Level-5 capabilities and to its AgentStack orchestrator in the second half of the year, and Siemens has said a self-verifying update to Fuse is forthcoming. Each is a chance for a named customer to appear outside a trial, which is the evidence still missing. ASP-DAC 2027 runs January 25 to 28 in Tokyo, where the HKUST overview is listed as an invited paper.

The agents converging on your synthesis tool right now are quieter than the ones trending on GitHub, and that is exactly why they are worth watching.

Sources:

- Toward Autonomous RTL and Synthesis Optimization: Benchmarking and Self-Improving Agents (https://arxiv.org/html/2611.13713)

- Dr. RTL: Autonomous Agentic RTL Optimization through Tool-Grounded Self-Improvement (https://arxiv.org/html/2604.14989v1)

- Dr. RTL repository, HKUST (https://github.com/hkust-zhiyao/Dr_RTL)

- RTLScout: Joint Agentic Code and Synthesis Optimization for Efficient Digital Circuits (https://arxiv.org/html/2606.06530)

- RTLScout repository, Huawei CSL (https://github.com/huawei-csl/RTLScout)

- ChipOmni: A General-Purpose AI Agent for Electronic Design Automation (https://arxiv.org/html/2611.13020)

- Can AI Agents Really Complete RTL-to-GDS? Lessons from Benchmarking Tool-Interactive EDA Workflows (https://arxiv.org/abs/2607.17528)

- Cadence Unleashes ChipStack AI Super Agent, Pioneering a New Frontier in Chip Design and Verification (https://www.businesswire.com/news/home/20260210926552/en/)

- Introducing Synopsys Long-Horizon AgentEngineer Solutions and Autopilot Platform (https://www.synopsys.com/blogs/chip-design/long-horizon-agentengineer-solutions.html)

- Synopsys unveils Autopilot platform for AI agents that run chip-design workflows (https://news.lavx.hu/article/synopsys-unveils-autopilot-platform-for-ai-agents-that-run-chip-design-workflows)

- Synopsys brings agentic AI workflows to chip design and verification (https://rcrtech.com/semiconductor-news/synopsys-agentic-ai-chip-design/)

- The state of agentic AI in chip design tools in 2026 (https://privacymatters.news/news/the-state-of-agentic-ai-in-chip-design-tools-in-2026-30602088598)

- ChipAgents Raises $74M to Scale an Agentic AI Platform to Accelerate Chip Design (https://www.businesswire.com/news/home/20260217568914/en/ChipAgents-Raises-%2474M-to-Scale-an-Agentic-AI-Platform-to-Accelerate-Chip-Design)

- Nvidia partner ChipAgents raises $60 million to accelerate chip design with AI agents (https://whbl.com/2026/07/29/nvidia-partner-chipagents-raises-60-million-to-accelerate-chip-design-with-ai-agents/)

- DeepSeek dsh-superagent-prototype repository (https://github.com/deepseek-ai/dsh-superagent-prototype)

- DeepSeek open-sources router-worker coding-agent prototype (https://aiweekly.co/alerts/deepseek-open-sources-router-worker-coding-agent-prototype)

- DeepSeek drops enhanced coding agent with aim to rival Anthropic (https://www.business-standard.com/technology/tech-news/deepseek-drops-enhanced-coding-agent-with-aim-to-rival-anthropic-126110100865_1.html)

- [arXiv digest] 2026-11-11: 7 in scope (+18 other) · Issue #19 · pact-safety/touchstone (https://github.com/pact-safety/touchstone/issues/19)

Disclaimer: This article is generated by GatiFlow Intelligence for informational purposes only. It does not constitute investment advice, recruitment recommendations, or legal guidance. All data is derived from public sources and AI analysis — verify independently before making decisions. Past trends do not guarantee future results.

Where this came from

Every Deep Dive starts from GatiFlow's own pipeline: 13 public developer sources, collected every six hours, with a confidence score and the evidence behind each signal. The same signals, filtered to the topics you follow, are a JSON API.

No credit card required.

Get the next one by email

One article every Saturday morning in your time zone. No account needed, and nothing else is sent to the address.

Double opt-in: you confirm by email first. What we store, and for how long, is in the privacy policy.

Tell me I am wrong

Corrections, the version of this you have lived through, or what you would like covered next. It reaches me directly and is never published. It is kept for two years so it can be read and answered; the privacy policy has the details.

0/2000