• conforms_to::[[Reference Form Contract]]
• serves_as::[[User-Space Pain Inventory Naming the Substrate Properties eOS Continuum Provides]]
• in_domain::[[eOS Continuum]]
• authored_by::[[Christopher Allen]]
• has_lifecycle::[[Seed Stage]]
• has_curation::[[Working Draft]]

Why I Built DSPy Agent Skills (codeandcontext.ai, 2025)

URL: https://codeandcontext.ai/why-i-built-dspy-agent-skills/

A blog post by the author of the dspy-agent-skills project arguing that DSPy research, tooling, and documentation exist separately, but there is no curated, spec-compliant way to hand an agent the relevant knowledge once. The post identifies that procedural knowledge differs fundamentally from raw information availability -- coding agents like Claude Code repeatedly fail to generate idiomatic DSPy code despite access to all the materials. Three building blocks named: (1) GEPA optimization where metrics explain why failures occur (not just return scalars); (2) RLM for long context, separating variable space (REPL memory) and token space (what the LM sees); (3) procedural workflow with a seven-step loop (spec then program then metric then baseline then GEPA then export then deploy).

Adopted

The post's "variable space vs token space" distinction is the [[Agent Harnesses Drive the Runtime, Not the Reverse|harness-as-tool inversion]] this graph names, restated in DSPy's vocabulary. The substrate holds state; the LM consumes a projection of it. The post also references "observational memory keeping agents from unauthorized state mutations" and "handoffs across agent instances" as related Agent Skills concerns -- both of which are exactly what [[Capability Boundaries Are Runtime-Enforced, Not Policy-Checked]] and [[Runtime State Is Persistent by Default, Not by Application Discipline]] provide as substrate primitives.

Not adopted (yet)

DSPy Agent Skills is a user-space framework on Python. The post acknowledges the gap (procedural knowledge is hard to embed; coding agents fail at idiomatic DSPy code) but solves it with embedded-knowledge skills rather than substrate primitives. eOS Continuum's substrate-LAYER position addresses the substrate gap that drives the user-space rebuild.

Sources

• URL: https://codeandcontext.ai/why-i-built-dspy-agent-skills/
• Date: 2025
• Venue: codeandcontext.ai blog
• Stub note: Authored 2026-05-03 as the public companion to the DSPy paper. Body to be expanded with per-building-block correspondence to substrate primitives (GEPA optimization to atomic operations plus state introspection; RLM long context to persistent state; procedural workflow to atomic operations across steps) in a future session.

Relations

• conforms_to::[[Reference Form Contract]]

  • Industry blog post serving as the public-facing companion to the DSPy paper; passes URL-resolvability.

• informs_downstream::[[Agent Runtimes Require Substrate Primitives, Not External Glue]]

  • The post identifies user-space pain points (variable-vs-token-space, observational memory, handoffs across instances) that are exactly the substrate properties this graph's diagnostic Conviction names.

• composes_with::[[DSPy Compiling Declarative Language Model Calls (Khattab et al., 2023)]]

  • The post is the public companion to the DSPy paper; both rest on the same Khattab-axis research program.

• composes_with::[[Recursive Language Models (Zhang et al., 2025)]]

  • The post explicitly integrates RLM as one of the three DSPy building blocks; the Khattab axis runs through both papers and this post.