Skip to content

Knowledge base

Architecting agentic systems

An agentic system splits into agents and workflows instead of services, and connects them through MCP, agent-to-agent protocols and hand-offs that carry context. Production use needs durable execution, tracing, bounded loops, trajectory evaluation and a zero-trust posture toward every agent.

Written by

Updated:

Agentic Engineering 5

Our position. We hold that the agent architecture that works is embarrassingly simple (opens in a new tab): a while loop, a few general-purpose tools and context managed through sub-agents, with no elaborate orchestration. Agent non-determinism is a problem architects already solved with eventual consistency (opens in a new tab): review critics, verifiers and tests are the compensating actions of a saga.

How do you design and operate a system built from agents?

The Die neue Architektur presentation sets the split into agents and workflows against domain-driven decomposition into services or modular monoliths, and gives each hand-off the context the next step needs, over MCP or agent-to-agent protocols (Die neue Architektur (opens in a new tab)). AWS’s pattern guidance anchors every agent design in three principles: asynchronous (agents operate in loosely coupled, event-rich environments), autonomy (they act without external control), and agency (they act with purpose toward specific goals), over a perception-reason-action loop (AWS - Agentic AI Patterns (opens in a new tab)). AWS draws the saga analogy differently from us. It reframes familiar distributed-systems constructs for reasoning: prompt chaining is the cognitive equivalent of sagas, where each thought is reevaluated, revised, or abandoned and rollback happens through reflective replanning rather than compensating transactions, and enrichment becomes dynamic and semantically guided instead of a static lookup (AWS - Agentic AI Patterns (opens in a new tab)).

Two composition styles are easy to conflate. AWS separates workflow orchestration, a central coordinator delegating a predefined sequence, from multi-agent collaboration, decentralized peers that negotiate and adapt through emergent coordination, and offers a catalog of roughly eleven reusable agent patterns (basic reasoning, RAG, tool-calling, tool-server, computer-use, coding, voice, workflow orchestration, memory-augmented, simulation, and observer) as composable blueprints (AWS - Agentic AI Patterns (opens in a new tab)). The case for specialization echoes biology: no single agent can do everything, and capable systems emerge from specialized agents composed together, the way the brain relies on specialized centers (Path to AGI: Beyond Words and Towards an Architecture of Reliable AI Reasoning (YouTube) (opens in a new tab)). That argument pulls against the minimal loop we start from.

Production agents need durable execution, not just good prompts. The presentation’s checklist separating real agents from demos names durable execution with checkpointing, full tracing (debugging without traces is hopeless), bounded loops with iteration limits, systematic evaluation with LLM-as-judge rather than only unit tests, context management, hierarchical agents with failure propagation, and parallel tool execution with partial state recovery (Die neue Architektur (opens in a new tab)). More agent loops are not free; more loops mean more context and more chaos (Die neue Architektur (opens in a new tab)).

Security gets its own posture: even internal agents are untrusted, with network and identity isolation per agent in a zero-trust, microsegmented model (Die neue Architektur (opens in a new tab)), and the hardest unsolved problem, prompt injection, cuts across all agent architectures equally, tool calling, MCP, and code execution alike, so no architectural choice makes it disappear (Markdown as Agentic UI (opens in a new tab)). Evaluation shifts from output checks to the trajectory: goal accuracy, tool-call accuracy, and whether the solution path was logical and efficient (Die neue Architektur (opens in a new tab)).

Products that embed agents get design rules of their own. Beyond parity and granularity, the Agent-native Architectures guide describes a composability flywheel in which the agent becomes a research instrument: rather than guessing features, you observe latent demand from what users ask the agent to do, and a failed request is itself signal about a missing tool (Every - Agent-native Architectures (opens in a new tab)). The audits are concrete: verify create, read, update, and delete exist for every entity, or the agent silently cannot fulfill obvious requests; keep tools atomic enough to work with types you did not know existed at build time; and when a hot path graduates to optimized code, keep the agent able to trigger it directly so parity still holds (Every - Agent-native Architectures (opens in a new tab)).

Autonomy for unsolicited actions is calibrated on a stakes-by-reversibility matrix, from auto-apply for low-stakes reversible actions to explicit approval for high-stakes irreversible ones, and because silent agents feel broken, every action must show in the UI and completion must be an explicit signal, never inferred by heuristic (Every - Agent-native Architectures (opens in a new tab)). A complementary interface insight: rather than invent a new agent protocol, match the runtime to the model’s training data, which is why markdown works so well as an agent-facing UI, arranging patterns the model already knows instead of teaching it new ones (Markdown as Agentic UI (opens in a new tab)).

Resources

Articles

Documentation

  • Agentic AI Patterns (opens in a new tab) - AWS prescriptive guidance cataloguing about eleven reusable agent patterns and separating workflow orchestration from decentralized multi-agent collaboration.

Posts

Talks and videos