2.9 KiB
2.9 KiB
agent design principles
Status: settled — as of 2026-07-16 Connects to: pipeline-stages, deterministic-gates, overview
Purpose
Fixes the shape every os-sdlc agent definition must follow: one job, minimal reads, minimal
tools, minimal returns. Load this before writing or reviewing any agents/*.md definition in
this plugin.
Design
- One clearly-defined job per agent. It reads only what it needs for that job; it returns the minimum under an explicit return format, so the next stage's context stays clean rather than accumulating everything every prior stage touched.
- Minimal tool allowlists via
tools:frontmatter inagents/*.md. The programmer/build agent getsRead/Write/Editonly — noBash. This is a deliberate[jrs]decision in the vault taxonomy note, stronger than IndyDevDan's "separate invocation" stance: rather than merely running build and test in separate agent invocations, the build agent is architecturally incapable of running tests or shells at all. NoWebFetch/MCP/other-plugin tools unless the specific job needs them. Effect observed in DeltaRefinery: agents scoped this tightly were more focused, cheaper per invocation, and faster — the constraint is a feature, not friction. - Settled facts travel in via the spawn prompt — an agent never "re-reads what the orchestrator already read." The orchestrating stage resolves context once and hands the agent exactly what it needs in the prompt itself.
- Inter-stage state travels via small on-disk handoff artifacts — a file-based analog of
DeltaRefinery's in-memory
Handoff(intra-level) and DB-persistedArtifact(inter-level). os-sdlc keeps this lightweight (files, not a database) because v1 is a single-session tracer-bullet pipeline, not DeltaRefinery's multi-session resumable system. - Model tiering: mechanical work → haiku; judgment work → sonnet; hard reasoning → opus. Applied per-agent-definition, not per-pipeline — a test-writer doing mechanical transcription of ticket acceptance criteria into test stubs is a different tier than a reviewer judging spec alignment.
Open questions
- Concrete handoff artifact file format (plain markdown, JSON, frontmatter'd markdown) — not yet chosen.
- Whether DeltaRefinery-style persisted resumability is needed before multi-session pipelines actually exist. Per the standing "don't build ahead of need" convention (see pipeline-stages's tracer-bullet framing), this stays undecided until a real multi-session use case shows up.
Sources
- SecondBrain vault:
agentic-sdlc-ai-developer-workflow-taxonomy.md([jrs]tool-allowlist stance vs.[dan]separate-invocation stance) ~/dev/delta-refinery(Handoff/Artifact pattern this generalizes from)plugins/os-sdlc/OVERVIEW.md- 2026-07-16 design session (this doc's origin)