Coco, a four-layer agentic platform, registers TPU co-design simulation sweeps into a SQL datastore and reports early deployment experience toward shorter time-to-simulation and time-to-insight

Coco is an agentic platform deployed with TPU architects that accelerates setting up experiments, sweeping simulators, and deriving insights in the co-design lifecycle. Whereas co-design analysis previously required manually working over hundreds of gigabytes of fresh simulation sweeps in heterogeneous files, Coco organizes this into a reusable agentic platform. The text describes early deployment experience 'deployed with TPU architects' and reports movement toward reduced time-to-simulation and time-to-insight, without giving specific quantitative values.

The platform uses a four-layer architecture: a datastore that automatically registers every simulation sweep into a normalized relational schema, a library of tools with typed APIs that agents compose, agents encoding recurring analysis workflows, and a platform UX whose navigation state doubles as agent context. Unlike naive 'chat-with-your-data' approaches, this design grounds every number an agent produces in a SQL query rather than scraping heterogeneous files. The layers are enumerated in the abstract as a system design description, without independent evaluation data for each layer.

The agents encode recurring analysis workflows, most notably iso-execution analysis, which compares systems at matched execution configurations and includes swept-but-dominated points off the Pareto frontier. By fixing the pattern of comparing systems at matched execution configurations into an agent workflow, dominated points also enter the comparison view rather than only frontier points. The abstract names the workflow as 'most notably iso-execution analysis' and states its coverage includes points off the Pareto frontier, without giving concrete case results for that analysis.

The authors argue that co-design is a distinct agentic domain: its data must be retrieved rather than memorized, its workflows are recurring but context-dependent, and expert adoption hinges on UX that balances IDE-style control with interactive exploration. This claim distinguishes co-design from general agentic application settings, noting that the evidence is by construction absent from pretraining corpora and there is no external literature to retrieve. This is a domain argument drawn from deployment experience, a position-level summary without accompanying controlled experiments or user-study data.

The work targets hardware-software co-design, especially analysis flows for jointly designing ML models and the accelerators that run them, and is deployed with TPU architects. Its applicability presupposes that evidence exists as simulation sweeps that can be registered into a normalized relational schema and grounded through SQL queries, and that analysis workflows are recurring but context-dependent. For simulation-heavy design teams facing the same situation where data is not in pretraining corpora and no external literature can be retrieved, this four-layer structure offers a reference organization; for expert users who need both IDE-style control and interactive exploration, the design in which navigation state doubles as agent context is a directly borrowable interaction idea.

The text is abstract-level and gives no specific reduction figures, measurement approach, or comparison baseline for time-to-simulation and time-to-insight, so the 'toward a reduction' phrasing cannot be further verified. How often iso-execution analysis is used in practice and with what effect, the independent contribution of each of the four layers, and the concrete relationship between expert adoption and the UX balance are not developed in the text. In addition, how the normalized relational schema covers heterogeneous simulator outputs, and the coverage and composition of the typed tool library, are open questions that require the full text to confirm.