AI is evolving fast. Production software needs engineering discipline.
AI agents as software components within an overall system.
AI agents as software components
We treat AI agents as software components within an overall system and bring the discipline of classic software development into agentic systems.
- Clear states and state transitions
- Defined interfaces
- Strict modularity
- A maintainable architecture
- Type safety
- Quality assurance and tests
Deterministic business logic remains where fixed rules and unambiguous processes apply, and is technically separated clearly from probabilistic components.
Deterministic logic
Fixed rules and unambiguous processes as classic code: the same result every time, fully testable.
Probabilistic components
AI agents for language, documents and judgment, embedded in a harness of guardrails, checks and evals.
Every AI concept has a classic counterpart
An AI agent is not a single model, but a software system that embeds a probabilistic model inside a controlled execution loop. We call this deterministic code around the model the harness.
The agent loop is a loop, human-in-the-loop an approval dialog, guardrails are validation, checkpointing is a snapshot, evals are regression tests. We have collected over 120 of these building blocks in nine areas.
Controlled AI agents
AI agents work within the framework the workflow sets and determine the next step themselves within it.
The harness gives them controlled access to data and systems, defined tools and only the permissions required for their respective task. They evaluate the results of their actions and can adjust their plan when needed. Actions and decisions are logged traceably and can be reviewed and limited within the workflow.
Guardrails and human-in-the-loop
AI acts within defined guardrails. The harness enforces them in code, not just in the prompt.
As soon as ambiguities, uncertainties or defined exceptions occur, the case is handed over to a human.
Human-in-the-loop is part of the system architecture and not just a fallback for errors.
Professional decisions, approvals and exception handling remain with humans where they cannot be automated reliably.
Testing, evals and observability
For probabilistic system decisions, classic software testing alone is not sufficient.
We therefore complement tests with automated evaluation frameworks (evals), defined test cases and ongoing observability.
This allows the quality of agents and workflows, as well as changes in live operation, to be checked continuously. Evals measure not the model alone, but model and harness together.
Production
For us, production includes reliable operation, monitoring, testing, maintenance and control of the overall system – not just the function of the individual agent.
Clear, modular and maintainable system architecture
We develop agentic AI systems with a clear architecture that is not unnecessarily dependent on a single model, provider or platform.
We therefore rely on
Clearly defined interfaces
Components talk to each other through fixed contracts and can be tested on their own.
Interchangeable components
Models, providers and services can be swapped without rebuilding the overall system.
Open standards
Open formats and protocols instead of proprietary platforms keep the system portable.
Containerized infrastructure
Every service runs in containers and behaves the same locally and in the cloud.
What matters is a clear, modular and maintainable system architecture – not being tied to a particular technology.
Our preferred tech stack
We choose the tech stack based on the requirements of the respective system, not the other way around.
The concrete stack can differ depending on requirements, the existing system landscape and architecture.
AI engineering
For AI engineering and agentic systems, we work primarily with Python and rely, among others, on Pydantic, LangChain/LangGraph and the Google Agent Development Kit (ADK).
- Python
- Pydantic
- LangChain/LangGraph
- Google Agent Development Kit (ADK)
Frontend & mobile
On the frontend, we develop TypeScript-first with React and also build the UI primitives for AI agents.
For web applications, we use Vite or Next.js depending on requirements, and for mobile iOS and Android apps, we use React Native and Expo.
- TypeScript
- React
- Vite
- Next.js
- iOS
- Android
- React Native
- Expo
Backend & infrastructure
We run backends and services containerized with Docker.
For relational data storage, we primarily rely on PostgreSQL, and for cloud infrastructure on the Google Cloud Platform (GCP).
- Docker
- PostgreSQL
- Google Cloud Platform (GCP)
Start a project or scale further?
Tell us where you stand and where you want to go. We support you along the way and will get back to you personally.
info@appamass.com