---
title: "Context design is a constrained tradeoff · Agentic Atlas"
description: "Context design is a constrained tradeoff: Context feasibility and diagnosability are designed jointly."
canonical: "https://agentic-atlas.dev/decisions/0009-context-design-objectives"
last-updated: "2026-08-19"
---

Manifest-selected Decision

Canonical id

0009-context-design-objectives

Status

accepted

Catalog order

3

Catalog revision `35263c4c415da742953d0462804fb14424e2244dae4c63efd27e468988de70ab`

# 0009: Context design is a constrained tradeoff

**Status:** accepted **Date:** 2026-07-14

## Context

The corpus needed a stronger stance on what "better" means. A simple ranking of correctness, observability, ease, and token efficiency failed in both directions: it could license throwing context at a workflow until it became impossible to run, or license deleting the evidence needed to understand a bad run. A rule to externalize anything mechanically possible was similarly local; wrappers can save tokens while increasing system-wide maintenance and failure cost.

## Decision

Context feasibility and diagnosability are designed jointly. Token efficiency is often a feasibility condition: a workflow that exceeds its usable window, dilutes load-bearing signal, or is economically unusable has not achieved its operational goal.

**Observability** means retaining the smallest structured evidence sufficient to reconstruct what governed a run, what context and tools influenced it, what effects occurred, which checks passed or failed, and where failure originated. It does not mean retaining every intermediate token. Receipts, manifests, assertions, identifiers, hashes, and compact event records are preferred over conversational archaeology.

There is no universal metric ranking. Patterns expose their forces and consequences so a workflow can be designed within its actual quality, context, latency, maintenance, and recovery constraints.

Prefer mechanizing work that can reasonably be scripted or shaped by a tool, including small CLI wrappers that reduce the LLM's interpretive burden. Apply a **net-system test**: authoring, discoverability, deployment, maintenance, and failure diagnosis count alongside token and variance reduction.

Burden relocation is a review question, not a mandatory template shape. Ask what improved, what worsened, and under which constraints; describe relocation only when a burden actually moved.

Deterministic orchestration remains an applicable pattern, not a default architecture. The higher-level question is which decisions genuinely require probabilistic judgment.

## Consequences

- Externalization is strongly preferred where the net-system test passes, but is not called an unconditional win.
- Context Efficiency owns the objective and evaluation lenses, not every normative statement in the family.
- Concept and pattern templates remain minimal; analytical lenses are applied in review rather than forced into every node.
- Efficiency claims must name the denomination being improved and must not hide lost diagnostic or quality capability.

[Browse all selected Decisions](https://agentic-atlas.dev/decisions)