$ today_in_10_seconds
At DAC 2026 this week, agentic AI crossed a threshold: from assisting engineers to running long, complex engineering jobs on its own. Synopsys unveiled fully autonomous chip-design workflows — a verification agent claiming up to 50x faster time-to-validated RTL — and Cadence and Siemens shipped parallel autonomy the same week. The recursive kicker: agents are now designing the chips that will run the next generation of agents.
$ lead_story
Synopsys pushes agentic AI from task automation to long-running autonomy
At the 2026 DAC Chips to Systems Conference, Synopsys unveiled fully autonomous agentic workflows for chip design powered by its AgentEngineer technology — explicitly extending agentic AI 'from task automation to long-running engineering execution.' The headline capabilities are concrete: a fully autonomous design-verification agent that orchestrates the entire verification cycle (claiming up to 50x faster time-to-validated RTL), the company's first fully autonomous CAE workflow for thermal analysis, and two new EDA workflows on Microsoft Discovery — actively evaluated by AMD — where an autonomous debug-closure flow already shows up to 40% cycle-time reduction. The distinction that matters for builders is 'long-running': these aren't single prompts but agents that coordinate domain-specific and task-level sub-agents across hours of multi-step engineering work. For anyone building agents, chip design is a proving ground with brutal feedback — the output either passes verification or it doesn't — which makes it a serious test of whether autonomy holds up on real, consequential work.
// why_it_matters: Agentic AI just moved from suggesting steps to owning entire multi-hour engineering workflows — on the silicon that will run the next generation of agents.
$ also_worth_attention
It's not just Synopsys — the whole EDA industry went autonomous this week
The bigger story at DAC is that all three EDA leaders crossed the same line at once. Cadence launched its AuraStack AI Super Agent for PCB and advanced packaging, completing a silicon-to-system agent portfolio, and Siemens added self-verifying agents to its Fuse EDA system — all built on NVIDIA's agentic stack (Nemotron, the OpenShell runtime, PhysicsNeMo). NVIDIA frames these as 'autonomous AI engineers working as digital coworkers' that compress weeks of simulation and verification into hours, with Nemotron 3 Ultra purpose-built for long-running agents (5x faster inference, ~30% lower cost). The insight: autonomous engineering isn't one vendor's bet — it's becoming the default architecture for an entire industry, standardized on a shared agentic runtime.
link →The capability that made this possible: task horizons are exploding
Long-running autonomy sounds like marketing until you look at the trend line. One widely-tracked benchmark measures the '50% time horizon' — the length of task, in human time, that an AI can complete with 50% reliability — and leading performance jumped from about 18 minutes to over two hours in a single year, with preliminary signs the same exponential holds in computer use and other domains. That's the quiet enabler behind this week's EDA news: agents can now sustain the multi-step, hours-long reasoning that real engineering requires. The insight: if the horizon keeps doubling, the set of jobs an agent can finish end-to-end expands fast — plan your roadmap around where the line will be in six months, not where it is today.
link →The new bottleneck: orchestration, not model smarts
As agents run autonomously for hours, the hard part shifts from the model to the system driving it — a practice developers are starting to call 'loop engineering.' As Claude Code's creator Boris Cherny put it, 'I don't prompt Claude anymore. I have loops that are running… they're the ones prompting Claude and figuring out what to do.' The emerging operating model across engineering teams is delegate, review, own: agents handle first-pass execution, humans review for correctness and risk, and accountability for architecture and trade-offs stays human. The insight: the scarce skill in late 2026 isn't prompt-crafting — it's designing the workflows, hand-offs, and review gates that let autonomy scale without diluting accountability.
link →
$ bigger_picture
The trend behind the headlines
This week is a genuine inflection: agentic AI graduated from task-assistant to long-running executor, and it happened across an entire industry at once. The through-line connecting today's stories is duration — the leap from an agent that suggests a step to one that owns a multi-hour verification or debug cycle — enabled by task horizons that have gone from minutes to hours in a year and standardized on shared runtimes like NVIDIA's. Chip design is the perfect first domain because the feedback is unforgiving and the payoff is recursive: agents are now designing the silicon that will run faster, cheaper agents, which will design better silicon still — a compounding loop at the base of the whole stack. But the same news relocates the human job. When agents run for hours unattended, value shifts from writing prompts to engineering the loops, hand-offs, and review gates around them, and accountability for the outcome stays firmly human. The teams that win the next phase won't be the ones with the smartest single agent — they'll be the ones who master orchestration and know exactly where a human has to sign off.
NVIDIA Nemotron (and NemoClaw/OpenShell)
Worth knowing because it's quietly becoming the shared substrate for autonomous engineering agents: Cadence, Dassault, Siemens, and Synopsys are all building on it. Nemotron 3 Ultra is a smaller, faster open model tuned specifically for long-running agents (NVIDIA cites 5x faster inference and up to 30% lower cost for complex agentic tasks), with the OpenShell runtime handling secure execution. If you're building agents that need to run for hours rather than seconds, it's a reference point for what a long-running-agent-optimized model and runtime look like.
$ subscribe --free
Get the next edition in your inbox
// DAIly Agent covers agentic AI three times a week
// free, no spam, exit anytime