Context as a Compiled View
Build each call's working context from richer durable state through named, ordered, observable processors, keeping storage separate from presentation, scoping by default and referencing artefacts by name rather than pasting them.
Evidence: moderate
In many harnesses the context window is the agent's state: messages are appended until something has to be cut, and whatever was cut is gone. That makes compaction irreversible, makes it hard to swap models, and leaves no clean record of what the model actually saw at each step.
Google's post on context engineering in its Agent Development Kit (ADK) inverts this. Its thesis is that context is a compiled view over a richer stateful system. Three principles follow. Separate storage from presentation, so the durable record is not the same object as what the model is shown. Make transformations explicit, as named, ordered processors whose effects can be observed. Scope by default, so sub-agents get the minimum context and fetch more through tools. State is held in tiers: the working context for one call, a session that is a durable log of typed events, memory, and artefacts that are referenced by name and version rather than pasted in. The post argues that typed event logs enable model swaps, compaction, time-travel debugging and observability. ADK itself filters events, summarises older turns, lazy-loads artefacts and tracks tokens automatically.
This is one vendor's architecture, but it agrees with Anthropic's Managed Agents design, where the session log sits outside the window and context transformation happens in the harness (see session log outside the context window and brain–hands–session decoupling). It is the agent-runtime version of the assembly step in prompt assembly algorithms, with the difference that the source is a durable event log rather than a fresh retrieval. It gives the four moves in write, select, compress, isolate a concrete pipeline shape, and it sits towards the structured end of the harness memory strategy spectrum.
Applied to an agent that investigates incidents, the pattern means keeping evidence (logs, health outputs, actions taken) in a durable store and building each model call's context from it through a few named steps, such as "last N health results", "log excerpt for component X" and "actions taken so far". Large artefacts are referenced by path rather than inlined. Reviewers can then see exactly what the model was shown when it made each decision.
Sources
- Google Developers Blog, 4 Dec 2025 — https://developers.googleblog.com/architecting-efficient-context-aware-multi-agent-framework-for-production/
- Google, ADK docs — https://adk.dev/