AMD has introduced Ross, an AI assistant designed to support developers in the areas of hardware, software and AI. What is special about it is less the assistant itself than the way it works: according to the manufacturer, Ross accesses AMD's own tools, documentation and stored engineering workflows. The use case lies in the development of embedded systems, i.e. software that is firmly integrated into devices and hardware.
Ross thus stands as an example of a development that is currently affecting many toolchains — the step from suggestion generator to agentic assistant.
From autocomplete to agent
The first generation of AI aids in software development worked essentially as extended autocomplete: it suggested the next line of code, completed function stubs or explained a section. The context usually remained limited to the open file.
Agentic assistants take a different approach. In this context, an agent is a system that not only outputs text but independently calls tools, checks intermediate results and executes several work steps in succession. Instead of an answer, a work result emerges — and with that, the question of what such a system actually needs in order to be useful also shifts.
The decisive raw material is called context
This is exactly where the revealing point of AMD's announcement lies. The three access points mentioned — tools, documentation, stored workflows — are not model properties. They are company assets.
A language model knows nothing on its own about a project's internal conventions: not which library version is mandatory, not why a particular interface looks the way it does and not otherwise, not which approval steps precede a deployment. This knowledge must be made accessible, otherwise the agent produces plausible-looking but project-alien results.
For companies, this leads to an uncomfortable insight: the benefit of an AI agent depends less on which model is purchased than on the state of one's own documentation and processes.
What this means in practice for SMEs
Very few organisations develop embedded systems. However, the structure of the problem is transferable. Anyone considering integrating agentic assistants into development or editorial processes should clarify a few points in advance:
- Is documentation available and up to date? Outdated instructions are worse for an agent than none at all, because it treats them as binding.
- Are workflows described explicitly? Much exists only as a tacit team convention. What is not written down cannot be stored.
- Which tools may the agent call? Read access to a repository is something different from triggering a deployment.
- How is it checked? Automated tests and reviews gain weight when more work steps are generated by machine.
Documentation becomes infrastructure
In many projects, documentation is regarded as a compulsory exercise that is budgeted tightly at the end. This calculation changes when documentation is no longer read only by people but serves as a knowledge base for assistance systems.
In practice, this means structured, versioned and machine-readable storage instead of scattered notes in emails, tickets and presentations. Decisions should record not only the what but also the why — rationales are more valuable to an agent than pure configuration lists, because they allow it to transfer to similar but not identical cases.
The same applies to processes. A workflow that exists only in the head of an experienced colleague is a risk — with or without AI. With AI, this risk becomes more visible because the agent starts guessing at exactly this point.
Realistic expectations
For all the dynamism, sobriety is worthwhile. The announcement describes an assistant for the support of development teams, not their replacement. Decisions about architecture, security requirements and priorities remain the task of people who know the business context.
Moreover: the more work steps a system executes independently, the more important control points become. An agent that independently writes code and operates tools needs clear boundaries, traceable logs and a review step before results reach productive systems. Anyone who does not plan these guardrails will trade typing effort for review effort — and gain little.
Assessment
For web and software projects, the lever thus shifts: it is not the choice of model that decides the benefit, but the quality of the context that an organisation can provide. Clean documentation, explicitly described processes and controlled tool access become a technical prerequisite rather than a diligence task. Anyone who invests in these foundations today creates the basis for using agentic assistants productively later without rebuilding — regardless of which tool prevails in the market.
Sources
- AMD lässt KI-Agenten Embedded-Systeme entwickeln — DataCenter-Insider | News | RSS-Feed