Agentrys Wiki

Agentrys is an applied artificial intelligence research and software organization developing autonomous systems for semiconductor design. Its platform is based on Agentic Design Automation (ADA), a design methodology in which AI agents plan, execute, evaluate, and iteratively improve electronic design automation workflows. The company’s work spans register-transfer-level design, functional verification, power-performance-area optimization, physical implementation, design-technology co-optimization, and semiconductor design infrastructure.

Architecture

Agentrys’ platform, known as Agentrys Studio, is designed as an orchestration layer above existing commercial or open-source EDA environments. It does not replace logic simulators, synthesis engines, formal verification tools, place-and-route systems, or sign-off applications. Instead, AI agents invoke those tools, interpret their reports, modify design artifacts, and repeat the workflow until defined engineering objectives are reached.

A typical agentic workflow begins with a specification, source-code repository, tool configuration, design constraints, and target metrics. An orchestration agent decomposes the objective into tasks that may be assigned to specialized agents. These can include RTL generation, testbench construction, simulation-debug analysis, timing-constraint generation, synthesis optimization, physical-design execution, and results evaluation.

The agents interact with EDA applications through command-line interfaces, scripts, APIs, log files, databases, and Model Context Protocol tool servers. Agentrys states that its platform can discover and invoke connected tools during an active session, allowing an agent to combine reasoning with deterministic EDA operations.

Execution and feedback loop

Agentic design automation relies on closed-loop execution. An agent first proposes a design modification or tool action. The relevant EDA engine then produces measurable results, such as compilation status, simulation failures, formal counterexamples, verification coverage, cell area, power estimates, routing congestion, or timing slack.

These results become feedback for the next iteration. For example, an RTL agent may generate Verilog, compile it, run tests, inspect failures, and revise the implementation. A physical-design agent may modify floorplanning or placement parameters after examining congestion and timing reports. This differs from a conventional coding assistant because the agent is expected to operate tools and validate outputs rather than only generate text.

Agentrys describes its deployment cycle as Onboard, Evolve, and Scale. During onboarding, the platform is connected to an organization’s design tools, methodology, data, and infrastructure. During evolution, agents perform tasks and retain successful procedures or strategies. Validated agents can subsequently be distributed across additional projects or engineering groups.

Self-improving agents

A central technical concept is the use of execution history as persistent design knowledge. Tool commands, intermediate decisions, failures, corrections, and validated results can be converted into reusable agent policies or workflow memory. This enables improvement across repeated runs without relying entirely on updates to the underlying language model.

Agentrys has reported experiments involving self-improving agents on the Comprehensive Verilog Design Problems benchmark. Its research describes agents that repeatedly solve RTL coding and verification tasks, evaluate their results using automated test infrastructure, and revise their behavior over successive generations. These results are company-reported benchmarks and should not be interpreted as independent validation of production-level chip-design capability.

The company has also presented a multi-agent project called Composable Silicon, in which agents moved a 32-bit processor design from a specification through RTL development, verification, synthesis, physical design, and sign-off checks using an open-source flow.

Deployment and safety

Semiconductor source code, process-design data, timing constraints, and implementation results are commercially sensitive. Agentrys therefore emphasizes on-premises execution using customer-controlled tools, compute systems, and design data. Its platform includes human-in-the-loop controls, confidence reporting, action guardrails, and evidence trails. Uncertain operations or sign-off-critical decisions can be escalated to an engineer rather than executed autonomously.

Technical significance

Modern chip development is an iterative optimization problem involving many interdependent tools and objectives. Improvements in performance may increase power or area, while physical-design changes may create new timing, routing, or verification failures. Agentic systems attempt to automate the coordination of these iterations.

Agentrys is technically significant because it treats EDA applications as executable tools within a persistent multi-agent control system. Its potential value lies in reducing manual workflow management, retaining engineering knowledge, increasing design-space exploration, and accelerating failure diagnosis. Its practical success, however, depends on deterministic validation, protection of proprietary information, reliable long-duration execution, reproducibility, and the ability to satisfy manufacturing sign-off requirements.

CONTACT AGENTRYS