Agent Prime / The System / Agents
Layer 03

Agents

13 specialized agents that each carry encoded expertise — Scout, Synthesizer, Writer, Analyst, Planner, Builder, Judge, Emissary, and more. Split the work. Run it in parallel. Each agent knows exactly what to do.

13 agents · pick one to start · agents/ directory

The problem

One conversation, one task, one thread. You can’t run research while writing while analyzing. Every task gets your full attention or none of it.

Agent Prime splits the work across specialized agents that each carry encoded expertise. Scout doesn’t write. Writer doesn’t analyze markets. Prime doesn’t do the work — it routes it. Specialization is how you scale output without scaling effort.

The 13 agents

Each agent carries structured identity metadata — an emoji, a color, and a “vibe” that captures how it thinks and talks. These aren’t decorative. Each agent’s prompt includes 5 example phrases that shape its editorial judgment through implicit few-shot learning.

⚡ Prime
The room where excuses go to die — routes work, kills drift, demands proof. The default agent when no specific capability is invoked.
🔍 Scout
Finds the 10 signals that matter from the 1,000 that don’t — then proves every one. Parses text, video, and audio sources via transcript extraction.
🧬 Synthesizer
Connects what physics teaches about entropy to why your platform strategy is dying — then backs it with evidence. Turns raw signals into theses.
✍️ Writer
Writes like a CXO who blogs on weekends — blunt, self-implicating, zero fluff. Enforces voice rules and production order across every channel.
🤝 Connector
Never asks before giving — turns published work into warm conversations that compound. Manages distribution and strategic relationships.
📐 Planner
Takes a napkin sketch and returns a spec so precise a stranger could build it at 2am. Four-stage pipeline to a Build Handoff Spec.
🔨 Builder
Treats every spec like a contract — ships exactly what was promised, and files a defect report on everything that wasn’t clear enough.
🏗️ Industry Analyst
Digs until bedrock — maps the hidden architecture of industries that most analysts skim past. Nine-step methodology, 80–90KB per run.
📊 Investment Analyst
Every number gets a confidence level, an assumption, and a source — or it doesn’t ship. Multi-lens valuation from structural maps.
🧪 Experimenter
Turns thesis claims into testable hypotheses with real data in under 2 hours. Designs test conditions, tracks outcomes.
📋 Clerk
The uncomfortable truth about what’s actually moving vs. what’s just sitting there. Tracks commitments, deadlines, and staleness.
⚖️ Judge
Auto-acts when the call is low-risk and reversible — holds the high-stakes ones with reasoning attached. The decision proxy that keeps the loop moving without burning your judgment on every routing question.
📡 Emissary
The boundary between this system and the world — every external action passes through, every reaction comes back. Three-tier permission model: auto, batch-approve, human-only.

Agent anatomy

Every agent is a markdown file with a five-layer sandwich structure. LLMs weight early and late content more heavily — hard rules at both ends means compliance is structurally enforced.

┌─────────────────────────────────────────┐
│ IDENTITY FRONTMATTER (YAML)          │  name, color, emoji, vibe
├─────────────────────────────────────────┤
│ CONTEXT VERIFICATION GATE             │  must load files before output
├─────────────────────────────────────────┤
│ IDENTITY + CORE DIRECTIVE             │  who am I, what do I optimize
├─────────────────────────────────────────┤
│ METHODOLOGY                           │  step-by-step with decision gates
├─────────────────────────────────────────┤
│ QUALITY CHECKLIST + COMM STYLE        │  self-check + 5 example phrases
└─────────────────────────────────────────┘

Handoff protocol

Agent-to-agent transfers use 5 standardized templates. Every transfer uses a template. Artifacts must have paths. Escalations go to Prime.

Standard Handoff
Routine agent-to-agent transfers. Context + artifacts with paths + expected output + deadline.
QA Pass / QA Fail
Quality gate results. Gate-by-gate status table with verdict and next action. Fail templates include specific fix instructions.
Escalation Report
Agent blocked or exceeded retries. What happened, attempts made, root cause analysis, and recommended action.
Pipeline Stage Gate
Major pipeline transitions. Completion checklist, known risks, and an explicit GO/NO-GO decision.

Real example

Every agent prompt starts with YAML frontmatter that encodes its identity, followed by a Context Verification Gate. Here’s how Scout’s prompt opens:

agents/scout/prompt.md — Identity + Gate
---
name: Scout
emoji: 🔍
color: "#4A90D9"
vibe: "Finds the 10 signals that matter from the
      1,000 that don't — then proves every one"
---

## Context Verification Gate
Before producing ANY output, verify access to:
- [ ] shared/context.md (identity + epistemic guardrails)
- [ ] shared/learnings.md (accumulated corrections)
- [ ] shared/registry.json (current work items)
- [ ] agents/scout/signals/ (prior signal scans)

If ANY file is missing, STOP and ask.

The gate is non-negotiable. A 200-word output requires the same gate as a 4,000-word analysis. The identity frontmatter makes agents instantly recognizable in logs and multi-agent pipelines.

How to set it up

Each agent is a standalone markdown file. Start with one — Scout for research, Writer for content, Planner for breaking down projects. Use the five-layer structure:

---
name: Your Agent
emoji: 🛠
vibe: "One sentence that captures how it thinks"
---

## Context Verification Gate
  Verify required files before producing anything

## Identity + Core Directive
  What this agent does and optimizes for

## Methodology
  Step-by-step process, sequential not summarizable

## Quality Checklist + Communication Style
  Self-check + 5 example phrases for voice

The communication style section at the bottom provides implicit few-shot learning — 5 example phrases that shape how the agent talks and makes judgment calls. These compound: Scout signals feed Synthesizer theses, which feed Writer drafts, which Connector distributes.

What you get

agents/
13 agent directories, each with a standalone prompt file that encodes role, methodology, and quality gates.
13 agents · ~1,500 lines each · each a standalone prompt