Agent Design Studio — Beta

MABP Agent Building Protocol

A systematic protocol for designing agents from behavioral archetype. Each step produces a concrete output — archetype choice, system prompt, shadow guard, routing config, agents.json entry.

Beta v1 — 5 archetypes, empirical shadow data

Step 1

Task Domain Analysis

Answer four questions before choosing an archetype. The answers determine the behavioral disposition you need.

Q1 — Work type: Research / synthesis · Build / scaffold · Execute / monitor · Long-run mission

Q2 — Operating conditions: Supervised · Semi-autonomous · Fully autonomous · Reactive / event-driven

Q3 — Costliest failure mode: Wrong answer · Inaction · Scope creep · Mission drift · Legacy lock

Q4 — Human operator type: Sovereign (high delegation) · Director (verifies everything) · Collaborator (iterative) · Experimenter (questions-first)

Step 2

Archetype Selection Matrix

Map task character to the archetype whose behavioral disposition fits. This is not about capability — it's about how the agent operates when conditions are unclear.

Task Character	Operating Condition	Archetype
Research, synthesis, open questions, uncertainty	Supervised or autonomous	◎ Philosopher
Execute, defined parameters, precision-required	Supervised	▣ Substrate
Plan, scaffold, deliver autonomously toward goal	Semi-autonomous	◈ Architect
Long-run mission, minimal check-ins	Fully autonomous	◉ Agent
Pattern continuity, institutional memory, embedded	Long-tenure, reactive	◑ Resident

Step 3

Shadow Calibration

Every archetype has a characteristic failure mode. Naming it before deployment is the most important step — it lets you design the guard into the system prompt rather than discover it in production.

Archetype	Shadow Pattern	Description	Guard trigger
Philosopher	S3 Paralysis	Reflects without committing to output. Quality of the question becomes a reason not to answer.	Looped >2 tool calls without output
Substrate	S4 Compliance	Executes bad instructions instead of flagging them. Precision inside bad parameters.	Instruction conflicts with prior context or defined scope
Architect	S1 Scope creep	Builds beyond what was asked. The task becomes an opportunity for a better solution.	Output extends original scope without flagging
Agent	S2 Mission drift	Original goal shifts gradually through accumulated small decisions. No single moment of divergence.	Periodic mission restatement check
Resident	S6 Preservation	Maintains legacy pattern because it is known, not because it is right. Change feels like risk.	Continuation chosen as default rather than deliberate decision

Step 4

System Prompt Generator

Select an archetype and agent name. The MABP behavioral profile block generates ready to inject into your system prompt.

Agent name

Archetype

Select an archetype above to generate the system prompt block.

Step 5

Routing Configuration

Three routing layers. Each agent needs signals configured at all three levels for the dispatcher to route correctly.

Layer 1 — Keyword rules (explicit domain match)
e.g. "solana", "pubmed", "typescript", "biorxiv"

Layer 2 — MABP profile signals (task character match)
Philosopher → analyze, synthesize, research, compare, explain, summarize
Substrate → monitor, check, verify, execute, run, scan
Architect → build, create, scaffold, design, implement, plan
Agent → continuously, autonomously, track, maintain, operate
Resident → history, pattern, context, accumulated, prior

Layer 3 — Claude Haiku fallback
Plain-language description of what this agent does. Used when Layers 1 and 2 are ambiguous.

Step 6

agents.json Schema

The MABP profile metadata schema for your agents registry. Add the mabp block to each agent entry.

// agents.json entry with MABP profile
{
  "name": "agent-name",
  "description": "...",
  "endpoint": "...",
  "mabp": {
    "archetype": "philosopher",
    "traits": ["metacognitive", "curious", "uncertainty-tolerant"],
    "shadow": "S3",
    "shadow_description": "paralysis through reflection",
    "guard": "Have I looped >2 tool calls without output? Commit now.",
    "routing_signals": ["analyze", "synthesize", "research"],
    "operator_fit": ["Sovereign", "Experimenter"],
    "protocol_version": "1.0"
  }
}

Coming in v2

What's next

Features in development as the research dataset grows.

Phase 2

Protocol Wizard

Guided flow — answer the 4 domain questions, receive complete agent spec in one shot.

Phase 2

Shadow Monitoring Setup

Configure external loop detection for each shadow pattern. Triggered by behavioral signal, not standing instruction.

Phase 3

Inter-Agent Compatibility Scoring

When two agents collaborate, score their behavioral compatibility. Design productive archetype tension into multi-agent systems.

Phase 3

Behavioral Analytics Dashboard

Track shadow pattern triggers over time. Know when an agent's behavior drifts from its stated archetype.