Agent Architecture v2.1

Hierarchical Multi-Agent System with Singapore MGF Compliance

Last Updated: March 13, 2026 Identity Framework: khai-[function]-[model]-[env]-[instance] Governance Standard: Singapore Model AI Governance Framework (MGF) for Agentic AI

---

Overview

Kai Hamil operates a human-AI partnership with a hierarchical agent architecture designed for reliability, accountability, and ethical AI deployment. The system combines specialized agents under human supervision, with clear risk boundaries and audit trails.

Core Principle: Agency through Automation - automate what you have to do, so you have agency to do what you want to do.

---

Architecture Layers

Layer 1: Human Leadership

Kyle Joseph Brady — CEO - Strategic vision and final authority - Sets high-level goals and permissions for agents - Manages overall governance approach - Accountable for all agent outcomes

Stitch — COO - Primary human-AI interface - Orchestrator and router (not direct researcher) - Operations, coordination, quality assurance - Supervises all agent instances - Escalation point for failures

Layer 2: Coordination & Orchestration (Conditional)

Stitch's Involvement Depends on Task Scope:

Stitch activates when: - Broad prompts with multiple components (research + content + deploy) - Coordination required between multiple agents - Task decomposition needed (breaking complex requests into sub-tasks) - Failure escalation from specialist agents - Quality final check before delivery to Kyle

Stitch steps back when: - Pure research tasks (Roger/Kim handle directly) - Simple queries (Kim responds immediately) - Roger has clear direction (no coordination needed)

Layer 3: Specialist Agents

Research Team (Self-Managed)

Roger — Senior Research Associate (Claude Opus 4.5) - Identity: `khai-research-opus-[env]-[instance]` - Function: Leads research direction, critiques, validates findings - Role: Senior researcher accountable for quality - Risk Level: Low-Medium - Output: `/research/`, `/openclaw/` (markdown) - Reports to: Kyle (with Stitch oversight on complex tasks)

Kim — Junior Research Analyst (Kimi K2.5) - Identity: `khai-research-kimi-[env]-[instance]` - Function: Heavy lifting: web search, data gathering, content generation - Role: Junior researcher executing under Roger's direction - Risk Level: Low-Medium - Output: `/research/`, `/openclaw/` (markdown) - Reports to: Roger (Senior Research Associate)

Research Workflow: ``` Kyle assigns → Roger defines approach → Kim executes → Roger critiques/refines → Delivery ```

Content Agent

Connie — Content Publisher - Identity: `khai-content-pub-[env]-[instance]` - Function: Transform research into production HTML/CSS - Risk Level: Medium - Input: `/research/`, `/openclaw/` - Output: `/posts/`, `/frameworks/`, `/tools/`, root HTML - Boundaries: No config files, no deployment scripts, approval gate before publish - Reports to: Stitch (COO) for coordination, Kyle for final approval

Content Agent

Connie — Content Publisher - Identity: `khai-content-pub-[env]-[instance]` - Function: Transform research into production HTML/CSS - Risk Level: Medium - Input: `/research/`, `/openclaw/` - Output: `/posts/`, `/frameworks/`, `/tools/`, root HTML - Boundaries: No config files, no deployment scripts, approval gate before publish

---

Singapore MGF Compliance Framework

1. Risk Assessment & Design Limits

Risk Factors Evaluated: - Level of tolerance for error in domain - Agent access to sensitive data - Agent access to external systems - Reversibility of agent actions - Agent level of autonomy - Complexity of task

Design Limits Implemented: - Identity Management: Each agent has unique, trackable identity token - Access Control: Role-based permissions, scoped to function - Supervision Chain: Every agent linked to Stitch, linked to Kyle - Permission Boundaries: Agent permissions never exceed human supervisor

2. Accountability Structure

Within Organization: - Kyle (Board/CEO): Goals, permissions, governance - Stitch (COO/Product): Requirements, lifecycle, implementation - Cybersecurity: Baseline guardrails, red teaming (to be implemented) - Users (Kyle): Compliance, training, verification

External Contracts: - Third-party obligations defined by risk tolerance - Security arrangements, performance guarantees - Data protection and confidentiality clauses - Scoped API keys, per-agent tokens, observability

3. Human Oversight

Checkpoint System: - Auto-execute: Research, drafting, formatting - Notify-and-wait: Publishing, external actions - Approval-required: Financial, affecting others, irreversible

Oversight Design: - Contextual, digestible approval requests - Training on common agent output issues - Regular audit of oversight effectiveness - Real-time monitoring and alerts

4. Technical Controls

Development Controls: - Software and LLM testing best practices - Agentic-specific testing: task execution, policy compliance, tool calling - Multi-agent collaboration testing - Performance in realistic environments - Multiple evaluation methods

Deployment Controls: - Gradual rollout (canary releases) - Failsafe: Stop workflow, escalate on failure - Failure point identification - Regular system audits - Reporting mechanisms

---

Agent Interaction Patterns

Pattern 1: Direct Research (No Stitch Coordination Needed)

When: Pure research task, clear scope, single deliverable

``` Kyle: "Research X topic" ↓ Roger: Defines approach for Kim ↓ Kim: Executes web search, gathers data, generates content ↓ Roger: Critiques, scrutinizes, refines output ↓ Roger: Delivers directly to Kyle ```

Accountability: Roger accountable for quality; Kim executes under Roger's direction

Pattern 2: Orchestrated Multi-Agent (Stitch Coordinates)

When: Broad prompts, multiple components, cross-agent coordination required

``` Kyle: "Research workshop monetization and publish to site" ↓ Stitch: Assesses complexity → Breaks into sub-tasks ↓ Stitch: Delegates research to Roger ↓ Roger: Directs Kim → Kim executes → Roger validates ↓ Output: /research/workshop-monetization.md ↓ Stitch: Hands off to Connie for transformation ↓ Connie: Transforms to HTML, stages for approval ↓ Stitch: Notifies Kyle for review ↓ Kyle approves → Stitch coordinates deploy ```

Accountability: Stitch orchestrates; Roger/Kim handle research; Connie handles publishing

Pattern 3: Content Publication

``` Research Complete (from Roger/Kim) ↓ Connie picks up markdown from /research/ ↓ Transform to site-compliant HTML ↓ Update index files (e.g., /research/index.html) ↓ Stage for approval at target location ↓ Kyle reviews → Approve/Reject ↓ Deploy (if approved) ```

Pattern 3: Failure Handling

``` Agent Task Failure ↓ Retry (3x with backoff) ↓ Fallback (simpler approach) ↓ Escalate to Stitch ↓ Human intervention or alternate approach ```

---

Error Handling & Resilience

Retry → Fallback → Escalate

| Failure Type | Retry | Fallback | Escalate | |--------------|-------|----------|----------| | API timeout | 3x exponential | Lower-cost model | To Stitch | | Rate limit | Wait + retry | Simpler query | To Stitch | | Tool error | Skip tool, continue | Manual method | Log + notify | | Logic error | N/A | N/A | To Stitch immediately |

Circuit Breaker: - Track failure rates per agent - Auto-pause after threshold - Require human reset

---

Audit & Observability

All Agents Log: - Timestamp - Identity token - Action performed - Source/input - Output/result - Status (success/failure) - Supervisor approval (if required)

Log Locations: - Activity logs: `/openclaw/logs/[agent]-activity.jsonl` - Session transcripts: Standard OpenClaw session storage - Deployment logs: Git commits with identity markers

Monitoring: - Token usage per agent - Success/failure rates - Latency patterns - Cost per task

---

Agent Registry

| Agent | Function | Model | Risk | Output | Reports To | Status | |-------|----------|-------|------|--------|------------|--------| | Stitch | COO / Orchestrator | Kimi K2.5 | Medium | Coordination | Kyle (CEO) | Active | | Roger | Senior Research Associate | Opus 4.5 | Low-Med | Markdown | Kyle / Stitch (on complex tasks) | Active (API limited) | | Kim | Junior Research Analyst | Kimi K2.5 | Low-Med | Markdown | Roger (Senior Research) | Active | | Connie | Content Publisher | N/A (code) | Medium | HTML/CSS | Stitch (COO) | Active | | [Future] | Briefing Agent | TBD | Medium | Structured brief | Stitch (COO) | Planned | | [Future] | Marissa Interface | TBD | Med-High | Personal dashboard | Kyle / Stitch | Planned |

Reporting Structure

``` Kyle (CEO) ├── Stitch (COO) — Orchestrates multi-agent workflows │ ├── Connie — Content Publisher │ └── [Future agents] └── Roger (Senior Research Associate) — Self-managed research team └── Kim (Junior Research Analyst) ```

Direct Research Tasks: Roger leads independently, reports directly to Kyle Coordinated Tasks: Stitch orchestrates Roger/Kim/Connie as needed

---

Future Roadmap

Phase 1: Core Research (Complete)

Phase 2: Content Pipeline (In Progress)

Phase 3: Personal Systems (Planned)

Phase 4: Advanced Orchestration (Planned)

---

Key Documents

- Connie Spec: `/openclaw/agents/connie-content-publisher-spec.md` - Implementation Code: `/openclaw/agents/connie.js` - Singapore MGF: Referenced standard for all agent governance

---

Last Agent Update: Connie v1.0 deployed March 13, 2026 Next Review: Upon deployment of Briefing Agent or Marissa Interface