Governing Artifacts

Summary

The artifacts that govern all the moving parts. In a human org these are “the playbooks”; in an agentic system they are the context layer — the ground truth agents read before acting. Weak artifacts = weak agents. This is the layer that must be built first.

The four artifacts in the original model

1. Brand guideline + brand identity + design system

Voice, tone, visual rules, components. Governs everything customer-facing.

• For agents: becomes the brand voice/policy context for any content task. See policies.md.

2. Raw footage + product mockups (+ asset library)

The reusable creative inventory — clips, photos, mockups, B-roll.

• For agents: the asset pool a content agent draws from instead of generating from nothing.

3. Market & competitor intel / knowledge

What the market wants, what competitors do, positioning, objections.

• For agents: this is the RAG knowledge base. Format it for retrieval (atomic notes, consistent frontmatter, dated). See Knowledge base as RAG.

4. Driver tree / value architecture

Decomposing revenue → actionable, controllable levers → KPIs for each moving part. The measurement spine of the whole system.

• For agents: defines what to optimize and which proxy metric to read. Without it, automation has no objective.

                         Revenue
            ┌───────────────┼───────────────┐
        Traffic        Conversion %        AOV
        ┌──┴──┐          ┌──┴──┐          ┌─┴─┐
     reach  freq    page  offer       price  bundle
       │                                  (controllable levers)
       ▼
   leading proxy (e.g. content "saves") ──► what an agent can actually move

The art is decomposing far enough that each leaf is controllable and has a measurable leading proxy. For content, you often can’t measure revenue per post, but you can measure saves/follows as a proxy that correlates with downstream conversion.

Scoreboard vs control panel (the key to setting agent targets)

An agent must optimize inputs, never lagging outcomes

Social isn’t performance marketing — you can’t set “gain 10k followers” as an agent target and have it act on that directly. The fix is the driver-tree discipline itself: decompose the lagging outcome down to controllable inputs, and let the agent act only on the inputs.

Follower growth        (lagging — you influence, can't control)
  └─ Reach / impressions served by algo   (lagging)
       └─ Retention + share rate + engagement velocity   (leading — controllable via craft)
            └─ Hook quality, format, topic fit, length, post time   (inputs — agent acts HERE)

So the agent never optimizes for follower count. It optimizes for retention rate, hook rate (% past 3s), completion, shares, and saves — because on discovery surfaces those are the inputs the algorithm converts into reach, and reach is what converts into followers.

• Scoreboard = lagging outcomes you watch (followers, reach, revenue).
• Control panel = leading inputs the agent steers (retention, shares, hook rate, saves).

Build a driver tree specific to each business model so the scoreboard→control-panel split is explicit. This is what turns “is saves a good proxy?” from a guess into a structured, testable decomposition. The analyze agent reads the control panel; humans watch the scoreboard. Applied in The proxy-metric insight.

The artifact the model is missing

Add a fifth: the Decision / Experiment Log

The reason this whole vault exists — “I forgot what I tested and have to relearn” — is solved by a lab notebook: every hypothesis, test, result, and learning, dated and tagged. It is the artifact that turns “learning while building” into compounding knowledge instead of repeated rediscovery. Template: decision-log.template. Argument: The decision log is your highest-leverage artifact.

Why artifacts come first (sequencing)

Build the context layer before the agents

The original realization is correct and worth stating as a rule: agents perform far better when the artifacts are set up first. They give the agent more context, more precision, and a better shot at a good output. Building the multi-agent pipeline before the artifacts exist means building agents that hallucinate your brand and guess your strategy.

Recommended order: brand + driver tree + knowledge base → one documented manual workflow → semi-automation with human gates → full automation of proven BAU. See 6. Crawl, walk, run — don’t build the full orchestra first.

Artifact → agent mapping (quick reference)

Artifact	Becomes (in agent terms)
Brand guideline	policies.md brand/voice rules
Asset library	The media pool an agent pulls from
Market/competitor intel	RAG knowledge base
Driver tree	The objective + proxy metrics an agent optimizes
Decision log	Memory of what’s been tried (anti-repeat)

Related

• Part A1 - Agent Operating Files — how artifacts become agent context
• Part A2 - Multi-Agent Orchestration — how agents consume them
• Three Entities — artifacts sit beneath all three