Roadmap: Context Engineering as strategic capability of the AI Harness

[dieterbaier]

Motivation

The current AI harness already provides reusable contracts, skills, adapters, metadata and deterministic generators. The next strategic step is to make context selection itself an explicit product capability.

Strategic direction

Evolve the toolkit from primarily Instructions-as-Code toward Context-as-Code.

The toolkit should not only define which rules and knowledge assets exist, but also make machine-readable:

• which context is required for a task;
• which context is optional or conditional;
• when context may be expanded;
• when an agent should stop searching;
• which steps should be deterministic instead of AI-driven;
• how resource use can be observed and constrained.

Proposed roadmap capability

Context Engineering

A toolkit-level capability covering:

• context resolution;
• context manifests for skills;
• metadata-first discovery;
• progressive context expansion;
• deterministic-first processing;
• context and execution budgets;
• stop conditions;
• reasoning profiles;
• sustainability reviews for AI workflows.

Possible delivery slices

Slice 1: Declarative context requirements

• Add required and conditional context declarations to skills.
• Update generated agent adapters to use those declarations.
• Avoid loading the full semantic contract by default.

Slice 2: Deterministic context resolution

• Implement a context resolver based on skill, metadata, relations, tags and changed files.
• Produce a minimal, explainable set of relevant files.

Slice 3: Progressive discovery and execution budgets

• Define expansion rules for documentation and source-code inspection.
• Add stop conditions and optional resource budgets.

Slice 4: Sustainability assessment

• Add review guidance or a dedicated skill for detecting unnecessary context loading, repeated AI work and tasks suitable for deterministic tooling.

Relationship to implementation issue

This roadmap capability provides the strategic umbrella for the concrete harness improvements described in #56.

Expected outcome

Context Engineering becomes a visible differentiator of the architecture-knowledge-toolkit: a clear separation of knowledge, selected context and agent execution, resulting in more focused, reproducible and resource-conscious AI-assisted architecture work.