AWS has introduced a runtime environment for AI agents: Strands Harness. It is open source, is intended to reduce token consumption and can be run locally. AWS thus addresses two points at which self-built agents regularly become expensive or unwieldy in practice – context management and the question of where the agent is actually executed.

What an agent runtime takes care of

An AI agent is a program that does not query a language model just once, but works in several steps: it plans, calls tools or interfaces, evaluates the results and decides on the next step. It is precisely this sequence that must be organised – and a runtime environment (in English runtime or harness) is responsible for this: it executes the loop, passes tool results back and records what has happened so far.

Anyone who builds this themselves quickly writes several hundred lines of infrastructure code that have nothing to do with the actual business logic. A ready-made runtime takes over this part and thus makes the application leaner and easier to maintain.

Context management as a cost factor

The context is everything a language model is supplied with in a call: instructions, conversation history, intermediate results, tool outputs. With multi-step agents, this context grows with every step. Since models are billed by token – small text or character units – the costs per run increase, and disproportionately so, because the complete history is transmitted again and again.

The announcement holds out the benefit precisely here: less manual context management in your own code and more economical use of tokens. How strong the effect is in concrete use cases can only be assessed seriously with your own measurements; comparative figures are not available at this point.

Local operation and data protection

The fact that the runtime can be executed locally is the most practically relevant point for many mid-sized projects. The orchestration – i.e. flow control, tool calls and state management – then remains in your own environment. This facilitates development, because agents can be tested without cloud access, and it creates scope for deciding which data leaves your own network at all.

The distinction is important: local operation of the runtime does not automatically mean that the language model also runs locally. If a model is accessed via an external programming interface, the transmitted content still reaches that destination. For a data protection assessment, it therefore remains decisive which models are connected and which content ends up in the context.

What open source means in practice

Open source code makes it possible to trace the sequence, connect your own tools and make adjustments without waiting for a vendor cycle. It also reduces dependence on a single platform. Conversely, responsibility shifts: updates, security issues and the maintenance of the integration lie with your own team.

Sensible checkpoints before use

  • Which licence applies, and does it fit the planned purpose?
  • Which models and providers can be connected – and is a change possible later without rebuilding?
  • How can the agents' behaviour be logged and traced, for example for troubleshooting and documentation obligations?
  • What costs actually arise per transaction, measured against a realistic use case of your own?
  • Who maintains the integration in the long term, and how great is the effort when versions change?

Assessment for project practice

The market for agent tools is currently moving quickly, and individual components are frequently replaced. It is therefore worth separating the business logic – i.e. rules, data access and approval steps – cleanly from the runtime. A later change then remains a manageable task instead of a rebuild.

For web and software projects, this means: an open source runtime that can run locally lowers the barrier to entry for agents in existing applications, for example for editorial support, data preparation or internal assistance functions. However, the benefit does not arise from the tool alone, but from clearly delimited use cases with a measurable result. A small, production-like prototype that demonstrates cost per transaction and result quality is recommended before a technology decision is rolled out more broadly.

Sources