What we are working on.

Working with agents

Specialists provide goals, drawings and references; agents support the work. We ask how context and feedback improve the quality of design variants, checks and drafts.

Questions and approach

• Prompting is evolving: goals, references and success criteria are gaining importance. For tasks such as drawing extraction or variant comparison, we aim to test which context helps and where additional rules make the work harder.
• Company knowledge should be available when needed. For us, context engineering means selecting the relevant information and keeping it current. Skills package recurring domain knowledge, instructions and tools for reuse.
• People set priorities, assess results and provide feedback. We examine which feedback should become better references or checks so that the next task needs less micromanagement.
• Recent Anthropic publications show how guidance changes as models improve. These approaches are to be tested against each process; shorter prompts are not a quality goal in themselves.

Building agent systems

Agent onboarding provides a concrete starting point: we set up agents for a specialist task with product knowledge, working instructions and suitable tools. We assess their value through the results.

Questions and approach

• We start with the business process: which data is reliable, which systems must be accessible and which actions are permitted? We clarify responsibilities, access rights and the steps requiring human approval.
• For known processes, we first consider fixed workflows. The same test tasks are to establish whether a single agent helps with variable inputs or clarification. Multiple agents become relevant when separate tasks or independent review can reasonably be expected to offer a measurable benefit.
• Repeated work cycles need a verifiable goal, stop conditions and limits on runtime and token use. Tokens are the text units processed by a language model. The next step must be clear when uncertainty arises or the budget is exhausted.
• The planned comparison uses representative test tasks, known as evals, to assess output quality, errors, duration and cost together. It is to examine actual outcomes in the target system and the steps requiring a human decision. A model change also calls for a fresh review of the system design.
• Neon illustrates an onboarding pattern: skills guide agents to relevant documentation and working practices, while connected tools enable actions. We apply this principle to specialist business tasks. The example demonstrates the technical approach; it does not yet establish benefits in our client processes.

Operating AI economically

Drawing checks, variant generation and quote preparation have different requirements. These inform the choice of models, operating environments and hardware. Result quality and total cost are the deciding factors.

Questions and approach

• The task, quality target, data requirements and workload determine the operating architecture. Our hypothesis is that many companies will operate such AI applications on an ongoing basis. Here, “AI factory” means the environment of data, models, tools and checks required to do so; scope and operating model must be determined for each application.
• Model architectures are to be compared against the task: dense models use their components throughout computation; mixture-of-experts models (MoE) activate selected subnetworks. Quantization stores model values at lower precision. Application tests must show whether this saves memory and cost while preserving the required quality.
• Model routing assigns tasks to suitable models. Caching reuses previously computed context, while batching groups requests. These methods are to be evaluated alongside response time and quality: the measure of token economics is cost per professionally accepted result, including failed attempts and rework.
• Hardware depends on the model and workload: a CPU as the main processor, a GPU or NPU as an accelerator, and working and model memory (RAM/VRAM). Memory bandwidth, networking and storage can limit throughput. Local, hybrid and cloud operations are to be compared against utilization, data requirements and operating effort.
• Total cost includes energy under load and at idle, cooling, depreciation, integration, services, maintenance, security testing, recovery from incidents, human review and rework. We investigate how consumption and shared infrastructure can be allocated transparently to departments and processes. Energy per result requires measurements on the actual system; a chip power rating is insufficient. Without measured data, cloud consumption must be labelled as an estimate.

Securing AI agents

An agent may be allowed to check a drawing without being allowed to send customer data or approve a quote. We ask how such boundaries can be enforced technically and verified within the professional workflow.

Questions and approach

• Minimal access rights, separate identities and narrowly scoped tools form the foundation. Permissions are enforced in the target system; critical changes and data transfers require verifiable approval. Our evaluation question: do these boundaries hold when assignments are incomplete or work is delegated to other agents?
• External documents, websites and tool responses can contain manipulative instructions, known as prompt injection. Such content must not grant additional authority. We examine data flows, permitted recipients and the origin of integrated tools and skills. Access keys belong in separate credential management, outside the model context. Local hosting alone does not ensure security.
• Manipulated inputs, attempted unauthorized actions and stop conditions are to be tested in isolated, authorized environments. We aim to record prevented boundary violations, legitimate work that is incorrectly blocked, and the effort required for human approvals. Actual actions in the target system are the evidence that matters.
• Operations require traceable logs without access keys, detection of unusual actions and consumption patterns, and a rehearsed way to stop execution, revoke access and recover. Responsibilities and the effort required for these controls are included in cost per accepted result.
• Dario Amodei’s September 2026 essay and Anthropic’s reports motivate these questions. We distinguish his assessment of future risks, observed misuse and unauthorized agent actions in misconfigured tests. The limitations of these sources are set out below; they do not replace evaluation of our specific applications.