A regression is the batch of tests that runs against the chip design every night, and RTL is the design written as code before it becomes silicon.
The nightly regression on the 7741-Y controller’s interface block finished at 05:10 with 312 failures out of 4,120 tests. By 07:30 Juno has grouped them into nine signatures. A signature is the pattern of the error message and the point in the test where it fails, so two failures with one cause carry one signature. Six signatures match open bugs and are attached to them with the log and waveform references; three are new and have gone to their block owners.
For the largest group Juno called the design tool’s own debug agent, which came back with a root cause in the interface arbiter and a proposed fix; the block owner has both in her queue, with the tool’s evidence beside them. The 240 tests that failed on a seed whose bug was marked fixed yesterday are rerunning.
The verification lead opens one page: coverage at 91.4%, two bugs past their service-level clock, three new signatures waiting for his confirmation. The 09:00 stand-up starts from that page. showcase estate
The design-tool vendors (EDA, electronic design automation) have moved fastest, and their figures are their own. Synopsys reports an autonomous debug closure workflow that cuts debug cycle time by 25% to 40%, with AMD evaluating it (Synopsys, 27 July 2026) 33. Its orchestrated design and verification workflow doubles productivity, up to five times in some cases (Synopsys, 11 March 2026) 34. Siemens ships a verification planning and debug toolkit used by MediaTek, and an agent endorsed by Samsung and NVIDIA with human checkpoints and audit trails (Siemens, 27 February 2026; Siemens, 16 March 2026) 41, 40; Samsung reports design turnaround down about 50% (Samsung, 17 March 2026) 43.
Cadence reports its ChipStack agent cutting verification effort by about ten times in some areas at Altera and formal verification time by up to four times at Tenstorrent. It produced a test plan in under 40 minutes against a week by hand (EE Times, 10 February 2026) 37. Its virtual engineer with NVIDIA claims a five-week validation loop done in under a day (Business Wire, 31 May 2026) 38.
The pain is measured. Verification takes 60% to 70% of a chip project’s engineering effort (Siemens, 3 September 2025) 32. First-silicon success fell to 14% in the 2024 Wilson Research Group study, the lowest in over twenty years, and 75% of projects ran behind schedule (Semiconductor Engineering, 27 March 2025) 31. McKinsey estimates that AI could nearly double the innovation rate in electronics (McKinsey, 20 June 2025) 49 and, in an earlier semiconductor study, cut R&D cost by 28% to 32% (McKinsey, 2 April 2021) 56.
Where the work sits is the honest part of this page. The design tools’ agents do the root cause and the fixes, inside the simulator and the debugger. Juno’s roles would run the flow around them: start and watch the regression, group and assign the failures, call the tool’s agent and record its answer. They would rerun what a fix should clear, keep coverage and bug convergence on one page and escalate when a clock runs out. The record of who decided what stays on the platform.
A nightly regression of thousands of tests runs through scripts and a job scheduler. Verification engineers spend the morning reading logs, grouping failures by hand, many of them duplicates, and assigning them to block owners by e-mail, chat and the bug tracker. Designers debug with waveforms and rerun; a fix that clears one seed is checked on the others by hand.
A weekly coverage review and a bug scrub meeting set priorities, and the programme manager builds status slides for the RTL freeze review from three tools. Ten to fifty engineers and six tools are involved, and the vendors put one verification loop at about five weeks (Business Wire, 31 May 2026) 38.
Illustrative times for a morning done by hand; the five-week loop is the vendor’s own figure 38.
Juno would sit beside the design tools, and the six stages below say what it would do at each. The regression manager, the coverage database and the bug tracker are registered and read in place; the design objects join the ontology; the triage policy is written in plain language; the design tools’ own agents are called as tools for the root cause; and every reassignment teaches the next run.
One Wednesday morning as the run trace would record it: the design tool’s agent is called at step 4 and answers at step 5; people decide at steps 7 and 8. Showcase estate; the roles are designed, yet to be built.
Inside the design tools the market’s level runs from execute within rules to autonomous: the three design-tool vendors describe fully autonomous debug closure and orchestrated workflows with engineer oversight, and say they are moving toward full autonomy (2026) 33, 38, 40. The blueprint designs Juno’s write-back roles here at A3: the agent executes triage, assignment and reruns inside versioned rules, and the verification lead handles exceptions only. Regression run orchestration and compute and licence scheduling are designed at A4, RTL freeze readiness at A2 with a named approver, and the freeze decision itself stays with people.
| Element | Setting | Set by | Changes |
|---|---|---|---|
| Objective | Every overnight failure grouped, owned and, where a fix exists, rerun before the 09:00 stand-up; coverage and open bugs in view toward RTL freeze | The verification lead, in a sentence | New version |
| Data scope | Regression results, coverage database, bug tracker, design data log and scheduler records for the 7741-Y programme | A. Novak, data lead, with the CAD team | New version |
| Rules | The triage policy and its planning table: a known signature to its last owner; a new signature on three seeds opens a bug; rerun on fixed | The verification lead with the Planner Agent | New version, simulated first |
| Trigger | On completion of the nightly regression, about 05:00, and of each continuous run; milestone dates; a bug’s clock; the freeze gate | The verification lead | At once |
| Actions permitted | Group and assign · open or update a bug · call the design tool’s agent · submit reruns · update the coverage tracker · publish the freeze report | The verification lead, owner | New version |
| Limits | Reruns per night 2,000 tests · compute above the daily farm budget waits for the CAD manager · one rerun per fix · grouping confidence floor 0.80 | The verification lead and the CAD manager | New version |
| Escalation | A bug past its clock, to the verification lead with its history · a new signature on a freeze-critical block, to the block owner at once · a failed write to the tracker, to the owner | The verification lead | At once |
| Level | A3 Delegated with the verification lead on exceptions; freeze readiness at A2 with the programme manager as approver | The engineering director | Up on evidence, down in one step |
| Feedback | Every reassignment, rejected root cause and waiver is recorded and used in the next cycle | Always on | Fixed |
Every row is a setting or a versioned rule; the level and the limits are the ones a manager signs.
Lima’s agent blueprint is a design of 960 agent roles across five end-to-end chains. It is the map of what Juno can be configured to do, and the first agents are built from it one at a time, the way the PO cancellation agent was.
Additional capability in Idea-to-Market lists the rest of the map for this chain.
| Measure |
|---|
| Debug cycle time, vendor-reported reduction of 25% to 40% (Synopsys, 2026) |
| Verification loop length, vendor-reported five weeks to under a day (Cadence, 2026) |
| Coverage, vendor-reported gain of 20% (Synopsys, 2026) |
| Failures triaged before the morning stand-up |
| Duplicate-failure share |
| Bugs open past their service-level clock |
| Time to RTL freeze |
| First-silicon success, baseline 14% (Wilson Research Group 2024 study) |
Baseline failures triaged before the morning stand-up first. It is counted from the tracker’s timestamps on day one, before any role is built, and it is the number the desk feels: how many of last night’s failures had an owner at 09:00. Debug cycle time and the verification loop length are the vendors’ headline figures and the table marks them as such. Time to RTL freeze and first-silicon success are read once per tapeout, the point where the finished design goes to the foundry.
The tools’ agents find the root cause and propose the fix inside the simulator and the debugger. Juno’s roles would run the regression, group and assign the failures, call those agents as tools, rerun what a fix should clear and keep the record.
Every new signature, every proposed root cause and every bug past its clock. Known signatures are assigned within the rule, and a rerun runs once per fix; the verification lead sees the exceptions in one queue each morning.
Where it is. The regression results, the coverage database and the bug tracker are read in place through the data fabric, with each engineer’s access rights applied, and only the triage records, bugs and rerun jobs are written back.
On evidence. The blueprint designs the triage roles at A3 and freeze readiness at A2. A level goes up when recorded cases pass the test and the owner signs; it comes down in one step, at once.
The regression found the bug and the block owner fixed it in the RTL. A fix that reaches silicon or a board becomes a change order. What does one change touch across the company?
Every assembly, plant, stock position and open order a change touches, found across the product lifecycle system and two ERPs, drafted for the change board.
Bring two systems and one decision your planners make every week. In the first week we register the data, confirm the mapping with your team, answer your questions in ChatLima and run the first agent at the level you choose.