Article Review: Group 7 — AI Agent Patterns

Articles Reviewed

1. Vibe Coding Agent Swarm with claude-flow — Adrian Cockcroft — Multi-agent orchestration using SPARC methodology, 150K+ lines in 2 days
2. "Services Are the New Software" — Han HELOIR — Analysis of Sequoia's autopilot thesis: sell the work, not the tool
3. Your AI Agent Is Failing Because of Context — Vivedha Elango — Context engineering guide: collection, storage, management, selection

Key Concepts

Multi-Agent Swarms (Cockcroft)

claude-flow orchestrates multiple Claude Code agents working in parallel: - SPARC methodology: Specification → Pseudocode → Architecture → Refinement → Completion - 5 agents with different roles (Core Architect, Specialist, Testing Lead, etc.) coordinate via shared memory - Key lesson: Start with planning/research phase before implementation — "just do the research, don't start implementation yet" - Incremental > monolithic: Asking for too much at once produces code that runs but doesn't match intent. Incremental iteration is far better. - Generated code is "ephemeral" — think of claude-flow as a "higher level idea-compiler" - Strong UX/design plans upfront produce dramatically better results

Intelligence vs Judgement Framework (Han HELOIR)

Evaluating which professional work AI can automate: - Intelligence: Rule-based, codifiable, complex but deterministic (contract clause comparison, medical coding, form filling) - Judgement: Requires experience, taste, instinct (strategic decisions, culture fit, novel situations)

The reliability gap: A copilot at 95% accuracy is useful (human catches errors). An autopilot at 95% accuracy is a lawsuit (no safety net). The gap from 95% to 99.9% is exponential in engineering effort.

Compound error cascade: At 95% per-step reliability, a 20-step workflow succeeds only 36% of the time (0.95^20 = 0.36). This is the lived experience of every team chaining autonomous AI steps.

Key insight: Companies cited as "autopilots" are actually hybrid systems — humans do load-bearing quality assurance. Design for hybrid from day one.

Context Engineering (Elango)

Context engineering is the systematic process of designing how you collect, store, manage, and use contextual information for AI agents.

Four eras: Structured input (1990s) → Natural language (2020) → Context-cooperative (emerging) → Machines understanding us (future)

Context processing strategies: - Timestamps — preserves order, no semantic structure - Role tagging — labels by function (goal, decision, action), improves retrieval - Q&A compression — fast lookup, loses holistic view - Hierarchical notes — tree structure, missing causal relationships

Memory management patterns: - Hierarchical memory — hot/warm/cold tiers (like CPU cache) - NL summaries — compress long context into natural language - Key fact extraction — fixed schema extraction from conversations - Semantic vectors — embed context for similarity search

Practical engineering tips: - KV caching for repeated context - Tool design that provides structured context - Context isolation between agents/sessions - Multi-agent context sharing via shared memory/state

Mapping to Our Architecture Repo & Claude Code Config

What We Do Right

1. SPARC-like workflow — Our rules/ai-assisted-development.md Rule 1 ("Architecture First, AI Second") mirrors SPARC's planning-before-implementation principle.
2. Incremental approach — Our strangler pattern (nge_enabled?) is inherently incremental — we don't rewrite modules wholesale.
3. Context management via CLAUDE.md — Our master CLAUDE.md provides structured context that loads automatically, similar to the role-tagging approach.
4. Skills as context injection — Our skills use progressive disclosure (Group 6 finding), which aligns with Elango's hierarchical memory pattern.

Improvements Identified

1. MEDIUM: Add Context Budget Guidance to Session Hygiene

CLAUDE.md has session hygiene tips but no guidance on context budget management. Elango's article makes clear that "too much context" is as harmful as "too little."

Add to CLAUDE.md session hygiene section: - Prefer focused reads (specific files) over broad searches (full directory scans) - Use Agent tool for exploration to protect main context window - When context grows large, summarize findings before continuing - Reset context (new session) when switching between unrelated tasks

2. LOW: Multi-Agent Orchestration Reference for ADR Implementation

Cockcroft's experience with claude-flow shows that multi-agent swarms can produce 150K+ lines but need strong upfront planning to match intent. For implementing ADRs 005-010, the SPARC methodology could be useful: 1. Specification — ADR defines what and why 2. Pseudocode — Design the event flow and module structure 3. Architecture — Map to existing patterns 4. Refinement — Review against rules 5. Completion — Implementation with testing

Recommendation: Reference SPARC in ai-assisted-development.md as an approach for large module implementations. Not as a tool dependency (claude-flow), but as a methodology.

3. INFO: Hybrid Human+AI Pattern Validates eDiscovery Review

Han HELOIR's analysis directly validates Nextpoint's document review model: - eDiscovery review is a hybrid process — AI assists (TAR, predictive coding) but attorneys make final decisions - The "intelligence" work (document classification, deduplication, near-duplicate detection) is what our NGE modules automate - The "judgement" work (relevance decisions, privilege calls, production decisions) stays with attorneys - Our nextpoint-ai module fits the "copilot" model — assists review, doesn't replace it

This framing is useful for positioning Nextpoint's AI features.

4. INFO: Context Isolation Pattern for Multi-Module Development

Elango's context isolation concept applies to how we work across modules. When developing in one module (e.g., documentloader), context from another module (e.g., documentexporter) pollutes the window.

Our Agent tool with subagent_type=Explore already provides isolation. The skills could be enhanced to explicitly scope context: "When exploring this module, do NOT load patterns unrelated to its type (Lambda vs ECS)."

Actionable Changes

Change	Target	Priority
Add context budget guidance to CLAUDE.md session hygiene	CLAUDE.md	MEDIUM
Reference SPARC methodology in ai-assisted-development.md	rules/ai-assisted-development.md	LOW
Add context scoping to explore-module skill	.claude/skills/explore-module/SKILL.md	LOW

Summary

The agent swarm article shows multi-agent orchestration is powerful but needs strong planning upfront — matching our "architecture first" rule. The services article provides a framework (intelligence vs judgement) that validates our hybrid AI approach in eDiscovery. The context engineering article reinforces that context management is the primary challenge, not model capability — supporting our CLAUDE.md + rules + skills approach. Main actionable improvement: add context budget guidance to session hygiene.

Sign In

Sign In

Please set a new password.

Set Password

Cancel

?

Ask the Architecture ×

Ask questions about Nextpoint architecture, patterns, rules, or any module. Powered by Claude Opus 4.6.

Ask