Synopsys ships long-horizon agents that run whole workflows

Most "AI agents" finish a task in under a minute. Synopsys just shipped a portfolio aimed at workflows that take weeks. The company announced AgentEngineer solutions and the Autopilot Platform today — long-horizon agents that reason, plan, and execute complete engineering workflows from silicon to systems. You are not buying EDA tools. But the architecture Synopsys published is the clearest public answer yet to a question every shop running agents hits: how do you keep an agent coherent across a job that does not fit in one context window?

What actually happened

Per Synopsys' announcement, the portfolio spans verification, system validation, implementation, analog, manufacturing, and simulation. Named workflows include autonomous coverage closure, software bring-up, multi-die 3DIC assembly, PPA closure, analog layout synthesis, mask synthesis, and signal integrity analysis.

The structure is three layers, and it is the part worth copying:

• AgentEngineers are domain-specific "super agents" that orchestrate other agents.
• Task agents complete specific, bounded tasks. They can be orchestrated by an AgentEngineer or called directly by a human.
• Engines — the actual tool layer — execute requested work but do not set goals or make decisions.

Synopsys says more than 50 customer engagements are underway, with general availability targeted for end of 2026. Partners quoted include Fujitsu, Intel, MediaTek, Nvidia, Samsung, TSMC, and AheadComputing. Fujitsu's Toshio Yoshida cites a 10–30% productivity boost in RTL code generation. Synopsys separately claims up to 50x faster verification closure, 20% higher coverage, and a 30% productivity boost — those headline figures are not attributed to a named customer in the release, so read them as vendor numbers.

Why long-horizon agents matter for your business

The three-layer split is the transferable idea. Orchestrator plans. Task agents do bounded work. Tools execute and decide nothing. Most agent systems we inherit have collapsed all three into one prompt that plans, acts, and calls the database, which is exactly why they drift halfway through a job and cannot be debugged afterward.

"Engines do not set goals" is a design rule, not a disclaimer. The moment your tool layer starts making judgment calls — a function that decides whether a refund is warranted instead of just issuing one — you lose the ability to audit where a decision came from. Keep the decision in the agent and the side effect in the tool.

Token efficiency is a stated design goal, which tells you the real constraint. Synopsys explicitly optimizes for token efficiency and latency. A workflow that runs for days burns budget in a way a chat completion does not. If you are pricing an agent that works a long queue, model the token cost per completed job before you model the headcount you save.

Task agents being directly callable is the escape hatch. An engineer can invoke one without the orchestrator. Build yours the same way: every step an agent can run autonomously should also be a thing a person can trigger by hand when the orchestrator makes a bad call at 2am.

Fifty pilots and a year-end GA is not a shipped product. Treat the customer quotes as directional. The one number attributed to a named company is Fujitsu's 10–30% on code generation — real, useful, and a long way from "autonomous engineering."

Sources: Synopsys Newsroom, Tom's Hardware.