Context Engineering

Definition

Structuring the information an AI agent loads — identity, task routing, stage contracts, reference material, working artifacts — so each step of a workflow sees only the context it needs, with every artifact a plain file a human can read and edit. Out-of-domain for the manipulator science; this topic exists because the project’s own agent harness (path-scoped rules, stage-scoped briefs, review gates) is an instance of it.

Key Equations

None — a methodology topic, not a physical phenomenon. Symbols per notation.md do not apply.

Source Support

• vanclief2026interpretable — Interpretable Context Methodology (ICM): folder hierarchy + markdown + local scripts in place of a coordination framework; five-layer context hierarchy; stage contract = Inputs/Process/Outputs. Its central quality claim (scoped context beats a monolithic prompt) is explicitly unmeasured — treat as hypothesis, not result.

Related Topics

• None in-corpus yet — the first methodology topic; aerospace topics do not link here.

Open Questions

• What would a fair scoped-vs-monolithic context benchmark on the same task look like? (Inherited from the source’s own threat-to-validity list.)

Vocabulary note

Tag context_engineering entered the closed vocabulary 2026-07-02 by user decision (see log.md), as the home for agent-methodology sources. agentic_methodology was proposed alongside and deliberately NOT added — one new tag per real need; revisit if a second methodology paper arrives.