Physical Intuition

Multiagent RL for chip design.

We train teams of agents on long-horizon RTL-to-GDS tasks. The main agent is the VP, orchestrating subagents and delivering the final handoff. Each specialist subagent gets its own tools, its own context, its own reward.

01 — System Design
01 — Single Agent

One agent. One long rollout.

Tries to do everything (architecture, RTL, DV, PD) in a single conversation.

User

design spec

→

Single Agent

one LLM

→

Long context rollout

> think about architecture...

> generate microarchitecture...

> write RTL...

> run simulation...

> debug errors...

> create testbench...

> write assertions...

> run formal...

> fix bugs...

> synthesize...

> place and route...

> analyze timing...

... (repeats)

~0K tokens

02 — Multiagent Collaboration

Specialists. In parallel.

Architect, RTL, DV, and PD agents collaborate through shared context and tools.

02 — Why Multiagent

Test-Time Compute Optimization

Chip design is simulation-heavy: every tool run dumps thousands of lines of logs, waveforms, and timing reports. Multiagent is how we allocate test-time compute per role — each specialist's full context window is spent on its own tool feedback instead of being diluted across the whole flow.

Reward shaping

A single agent gets one sparse reward at tapeout. Separate agents get dense, role-appropriate rewards: MicroArch is rewarded on performance against the golden model, DV on coverage and bugs found, PD on timing closure and DRC, RTL on passing checks.

Case Study: Multiagent Reinforcement Learning for Design Verification

A VP of DV main agent orchestrates subagents that each own one verification leg: formal, UVM testbenches, simulation, emulation. The VP reads the verification plan, assigns coverage targets, and triages failures back to the right subagent.

A single agent gets context polluted across methods, no clean per-method reward, and failures triaged by the same polluted context that caused them.

VP of DV

main agent

↓ tasks + coverage targets · ↑ rewards

Formal

proofs, cex

UVM

coverage, bugs

Simulation

pass rate

Emulation

bugs, cycles

session 1 · main agent

session 2 · subagent (role X)

session server · one model

TITO · clean per-role trajectories → data buffer

trainer · one update, mixed batches

knob: batch composition + per-role advantage scale

orchestration · puppeteer between sessionsreward hook · joint episode → both streamsone policy · no weight sync · both roles improvewatch: gradient interference

before — step 0 · GLM-4.6-Air base

> one 800-line testbench covering everything

> mixes formal + sim in one file

> no coverage targets assigned

after — step 70

> dispatches Formal on assert_*

> UVM coverage targets: 32 bins, 95% goal

> triages cex → Simulation subagent

0.00.30.60.90.000.040.080255070mean rewardKLsteps

GRPO on top of GLM-4.6-Air, illustrative

03 — Team

Production RL experience from xAI, OpenAI, and Ricursive, with engineers from NVIDIA and Fireworks.

Our team has trained frontier models, built large-scale inference stacks, and taped out real silicon. We're passionate about using model intelligence to build the best infrastructure.

04 — Selected Research