NDC London 2026

#conference

Conference Notes

Took notes with Claude and this summary was generated by Claude

Dates: 28-30 January 2026
Role: Volunteer

Table of Contents

Talks Attended

1. Beyond Pub/Sub: Advanced Messaging Patterns

Speaker: Poornima Nayar
Time: Wednesday 11:40-12:40, Room 2
Link: https://ndclondon.com/agenda/beyond-pubsub-advanced-messaging-patterns-07od/0opisnwzfkt

Key Patterns Documented:

Outbox Pattern
Ensures atomicity between database writes and message publishing.
Resource: https://microservices.io/patterns/data/transactional-outbox.html

Saga Pattern
Coordinates long-running workflows (minutes to days). Two variants:

AWS Step Functions implements orchestration variant.
Resources:

Scatter-Gather Pattern
Parallelise work, fan-out/collect replies, short-lived.
Resource: https://www.enterpriseintegrationpatterns.com/patterns/messaging/BroadcastAggregate.html

Competing Consumers
Solves throughput pressure.
Resource: https://learn.microsoft.com/en-us/azure/architecture/patterns/competing-consumers

Retry Strategies

Error Queue Management
Dead Letter Queues (DLQs) enable investigation/recovery.
Resource: https://docs.aws.amazon.com/AWSSimpleQueueService/latest/SQSDeveloperGuide/sqs-dead-letter-queues.html

Pinball Architecture
Mentioned (details pending)

Pattern Selection Guidance:

Relevance: Directly applicable to public sector operational resilience. Experience with AWS Step Functions validated understanding of Saga orchestration pattern.

2. Anatomy of an Incident (featuring CrowdStrike)

Speaker: Liam Westley (Head of Engineering at FreemarketFX fintech)
Time: Wednesday 11:40-12:40, Room 2
Link: https://ndclondon.com/agenda/anatomy-of-an-incident-featuring-crowdstrike-0jpk/2b70d36fec8b
Tags: Security, People, Soft Skills

Session Overview:

Builds on speaker's NDC London 2022 talk on incident handling. July 2024 CrowdStrike incident tested their 3-year refined incident plan.

Covers:

Career includes: chellomedia, GMTV, BSkyB, SmashedAtom.

3. The New Frontend Toolkit: Popover API, Dialog, and Next-Gen CSS

Speaker: Eric Veliyulin (Frontend specialist at Forte)
Time: Wednesday 13:40-14:40, Room 5
Link: https://ndclondon.com/agenda/the-new-frontend-toolkit-popover-api-dialog-and-next-gen-css-0vq5/0svlcd5xmx7
Tags: Web, JavaScript, UI

Problem Statement:

Developers still use third-party libraries for menus/modals/overlays, but native solutions now exist via Popover API and Dialog element. Almost no one using them.

Session Content:

4. Usability Testing is Easier Than You Think

Speaker: Jo Minney (Small business founder, technical communicator, Perth WA)
Time: Wednesday 15:00-16:00, Room 5
Link: https://ndclondon.com/agenda/usability-testing-is-easier-than-you-think-0hio/021t3vm4l5k
Tags: Testing, UX

Key Points:

Speaker Background:

5. Prompt Failures and Latency Spikes: Observability for AI

Speaker: Prerit Munjal (CTO, KubeCloud)
Time: Wednesday 16:20-17:20, Room 5
Link: https://ndclondon.com/agenda/prompt-failures-and-latency-spikes-observability-for-ai-0ud4/0zbj35ju4jr
Tags: AI, Cloud, DevOps, Tools

Problem Statement:

Traditional logs/metrics insufficient for debugging AI agents with unexpected responses. AI observability requires understanding prompts, latencies, retries, hallucinations, token usage, model behaviour under pressure.

Session Content - OpenLit:

Lightweight CNCF tool for LLM/agent workloads:

Use cases:

Accessible: No PhD in tracing or LLM internals required.

Speaker Background:

CTO of KubeCloud (umbrella company, multiple Cloud-Native products). Designs resilient architectures that scale while prioritising cost, security, availability, end-user experience.

Relevance: Directly applicable to IBM RAG and Agentic AI Professional Certificate work and AI development experimentation. OpenLit worth exploring for AI projects.

6. AI Agents Need Permission Slips

Speaker: Heather Downing
Time: Thursday 16:20-17:20, Room 2
Link: https://ndclondon.com/agenda/ai-agents-need-permission-slips-0hm0/0gk6h5107zb
Tags: Security, AI, Architecture, DevOps

Problem Statement:

MCP servers connect AI agents to enterprise systems, but examples lack specific access permissions. Works until assistant reorganises file system or emails entire customer database.

Core Message: "Giving AI agents broad permissions = giving toddler car keys."

Key Content:

ABAC Evolution (Attribute-Based Access Control):

Decision based on:

Note: "Context changes everything - same request at 3 AM from coffee shop? DENY"

The Permission Slip Framework: 5 Steps to Deploy Agents Safely

  1. Start with Least Privilege

    • Deny by default
    • Allowlist minimally
    • Scope to resources

  2. Implement Approval Gates

    • Identify checkpoints
    • Use interrupts
    • Human = ultimate arbiter

  3. Use Modern Authorization

    • OAuth 2.1 + scopes
    • ABAC for context
    • External policy engines

  4. Build Observability In

    • Audit trail everything
    • Explainability
    • Reversibility

  5. Separate Environments

    • Dev/staging/prod
    • Automatic separation
    • Test in sandbox

Speaker Background:

Passionate coder and entrepreneur. Experience: .NET, mobile apps, voice interfaces, IoT, databases. Focus: Coding for humans, right tool for job. Fortune 500 experience building enterprise mobile/.NET applications. Conference speaker supporting new developers. Competitive equestrian learning mounted archery.

Relevance: ABAC model significantly more sophisticated than RBAC - considers full context (who/what/when/where/why). Critical for AI agent security in production.

7. Meditations on Code As Art

Speaker: David Whitney
Time: Thursday 17:40-18:40, Room 3
Link: https://ndclondon.com/agenda/meditations-on-code-as-art-0ylg/0vuo674cdt8
Tags: Architecture, Ethics, Fun, People

Session Theme:

Can code deliberately be artistic? Despite how important and prolific software has become, the general public still have loose understanding of what software is. This hasn't been lost on developers, who try to capture themselves in the work they do.

Session Explores:

Notes from Session:

Apollo 11 Example: David mentioned that the function name "burn baby burn" was in the software for the 1969 Apollo mission. Discussed possible origin of the phrase in the Apollo codebase.

Key Theme: Fascinating intersection of software craftsmanship, history, and human expression in code. Developers leave traces of themselves and their culture in software, from NASA missions to modern applications.

8. How I Tamed Claude

Speaker: Rendle (CTO of XRAI)
Time: Friday 13:40-14:40, Room 1
Link: https://ndclondon.com/agenda/how-i-tamed-claude-0bb7/0fn2zo3kfbo
Tags: AI, Architecture, Machine Learning, Tools, Work skills, Cross-Platform

Speaker Background:

CTO of XRAI, on mission to subtitle the world and help people communicate with "magic glasses". Uses skills from professional stand-up comedy to deliver entertaining and informative talks. Recently learned bass to join tech parody band The LineBreakers.

Context Note: Talk submitted 6 months ago, had to change many times due to rapid rate of change in AI coding tools.

The Complete Four Personas Workflow:

Rendle's workflow involves creating 4 specialized personas for Claude using Claude Desktop to craft system prompts for each role:

  1. Analyst: "Gather and clarify requirements"

    • Asks questions to understand what's needed
    • Produces REQUIREMENTS.md document

  2. Architect: "Specification and Implementation plan"

    • Takes requirements document
    • Designs solution architecture
    • Creates detailed implementation plan
    • Produces architecture specs and technical plan

  3. Developer: "Execute the plan"

    • Takes architecture and implementation plan
    • Writes actual code
    • Implements solution according to specifications
    • Produces working code

  4. Reviewer: "Check generated code"

    • Reviews code created by Developer
    • Validates quality/correctness/security
    • Provides feedback for improvements
    • Produces code review feedback and approval

The Complete Pipeline:

Analyst → REQUIREMENTS.md → Architect → SPECS.md/PLAN.md → Developer → CODE → Reviewer → FEEDBACK

Workflow Advantages:

System Prompt Creation Process:

Uses Claude Desktop to meta-prompt: "Create a system prompt for a [Persona] that [specific role description]"

Comparison to Traditional SDLC:

This mirrors the traditional Software Development Life Cycle but with AI agents:

Key Innovation: Instead of trying to make one agent do everything (which leads to context overload and quality issues), Rendle creates a specialized agent pipeline that mimics proven human software development patterns.

Application: This is brilliant - applying Conway's Law in reverse. Instead of letting org structure dictate system architecture, structuring AI agents according to natural boundaries of software development process. Each agent has clear interface (document format) and responsibility, making system more maintainable, debuggable, and reliable. Perfect for government work where auditability, clear ownership, and ability to explain what happened at each stage are essential.

9. Resilient by Design

Speaker: Chris Ayers (Senior Software Engineer at Microsoft)
Time: Friday 15:00-16:00, Room 5
Link: https://ndclondon.com/agenda/resilient-by-design-0nqr/1d48800e63da
Tags: Architecture, Cloud, Microservices, Microsoft, Testing, Tools, Web

Session Description:

A truly resilient Azure infrastructure does more than withstand disruptions—it keeps applications running smoothly, no matter what. Remember: failure is always an option, but there are ways to reduce its impact.

Session Coverage:

Core Principles of Resilience:

Azure's Resilience Toolkit:

Real-World Strategies:

Speaker Background:

Chris Ayers is a Senior Software Engineer at Microsoft, where he builds AI-driven tooling to analyse outages and incidents across Azure. His work focuses on improving the way engineers understand failures, uncover risks, and strengthen reliability at cloud scale.

Background spans DevOps, cloud architecture, and modern application design. He's a frequent international speaker who shares real-world lessons about developer productivity, reliability, containers, and automation. Outside of work: proud dad, lifelong gamer, avid reader, and unapologetic nerd.

Team: Azure CXP (Customer Experience Platform) AzRel (Azure Reliability)

Tools/Resources Mentioned:

uptime.is - https://uptime.is
Tool for calculating what your uptime percentage actually means. Helps translate availability percentages into real downtime:

Why it matters: Helps teams understand the true cost/effort required to achieve different availability targets. The jump from 99% to 99.99% is not just a couple of nines—it's a massive engineering effort.

Azure Well-Architected Framework
Microsoft's framework (similar to AWS Well-Architected Framework). Six pillars:

  1. Operational Excellence
  2. Security
  3. Reliability
  4. Performance Efficiency
  5. Cost Optimization
  6. Sustainability (recently added)

Note: AWS created the original Well-Architected Framework in 2014; Azure followed with their version. Core five pillars are nearly identical across both cloud providers. The principles are largely cloud-agnostic.

Chris's Additional Resources:

Azure Architecture Centre
https://learn.microsoft.com/en-us/azure/architecture/

Microsoft's official catalog of solution ideas, example workloads, and reference architectures. Contains:

Azure Proactive Resiliency Library (APRL)
https://aka.ms/aprl

Curated catalog of resiliency recommendations for Azure workloads. Contains:

How it works:

  1. Run collector tool against Azure subscription
  2. Executes ARG queries to find non-compliant resources
  3. Get Excel spreadsheet showing issues
  4. Generate PowerPoint reports for stakeholders

Azure Verified Modules (AVM)
https://aka.ms/AVM

Microsoft's official, standardised Infrastructure-as-Code (IaC) modules. Available in:

Module types:

  1. Resource Modules: Single Azure resources with best practices
  2. Pattern Modules: Multiple resources for common architectures
  3. Utility Modules: Reusable functions/routines

Benefits:

Azure Review Checklists
Additional tool/resource mentioned. Likely checklists for manually reviewing Azure architecture against Well-Architected Framework principles.

Enterprise Web Apps
Resource mentioned on final slide.

How They Work Together:

  1. Azure Architecture Centre → Shows you what to build (reference architectures)
  2. APRL → Checks if what you built is resilient (scanning + recommendations)
  3. Azure Verified Modules → Gives you how to build it properly from the start (IaC templates)

Ideal Workflow:

  1. Find your architecture pattern in Azure Architecture Centre
  2. Implement it using Azure Verified Modules (already resilient by design)
  3. Run APRL scans to validate and find any gaps
  4. Fix issues and maintain over time

Contact Information:

Key Tools & Resources

Messaging & Architecture Patterns

AI Observability

Azure Resources

Well-Architected Frameworks

Azure Well-Architected Framework (6 pillars):

  1. Operational Excellence
  2. Security
  3. Reliability
  4. Performance Efficiency
  5. Cost Optimization
  6. Sustainability

AWS Well-Architected Framework (6 pillars):

  1. Operational Excellence
  2. Security
  3. Reliability
  4. Performance Efficiency
  5. Cost Optimization
  6. Sustainability

Note: Core five pillars are nearly identical. AWS created the original in 2014; Azure followed. Principles are largely cloud-agnostic - implementation differs (AWS vs Azure services).

MCP Server Recommendations

From Rendle's "How I Tamed Claude" Talk

1. Serena MCP Server

What it is: Free, open-source AI coding agent toolkit providing semantic code understanding through Language Server Protocol (LSP) integration.

Installation:

claude mcp add serena -- uvx --from git+https://github.com/oraios/serena serena start-mcp-server --context ide-assistant --project $(pwd)

Why it's recommended:

Key Features:

How it works:

# Instead of: "Search for 'authentication' in all files"
# You ask: "Where do we handle user authentication?"
# Serena understands semantic meaning and finds all related code

Use Cases:

GitHub: https://github.com/oraios/serena
Documentation: https://oraios.github.io/serena/

2. Context7 MCP Server

What it is: MCP server providing real-time, version-specific documentation for libraries and frameworks, injected directly into Claude's context.

Installation:

claude mcp add context7 -- npx -y @upstash/context7-mcp@latest

Why it's recommended:

Key Features:

How it works:

# Add "use context7" to your prompts
> Create a Next.js middleware that checks for JWT. use context7

# Or better: Set up auto-invoke in CLAUDE.md
> Always use Context7 MCP when I need library/API documentation

# Then just ask naturally:
> How do I set up Prisma with PostgreSQL?
# Context7 fetches current Prisma docs automatically

Specific library syntax:

# Tell it exactly which library to use:
> Implement authentication with Supabase. use library /supabase/supabase

# Request specific version:
> How do I set up Next.js 14 middleware?

GitHub: https://github.com/upstash/context7
Website: https://context7.com

Why These Two Together?

Rendle's recommendation of both Serena and Context7 creates a powerful combination:

Problem Without MCPs With Serena + Context7
"Where is authentication handled?" Greps through code, finds text matches Serena finds semantic relationships
"How do I use FastAPI auth?" Claude hallucinates old API Context7 fetches current docs
"Refactor the user service" Works on limited context Serena understands full symbol relationships
"Implement Stripe webhooks" Outdated code examples Context7 gets current Stripe API 
# 1. Install Serena (semantic code understanding)
claude mcp add serena -- uvx --from git+https://github.com/oraios/serena serena start-mcp-server --context ide-assistant --project $(pwd)

# 2. Install Context7 (current documentation)
claude mcp add context7 -- npx -y @upstash/context7-mcp@latest

# 3. Create a slash command to tie it together
# .claude/commands/go.md:
---
name: go
description: Start working on a task with full context
---

Always use:
- serena for semantic code retrieval and editing
- context7 for up-to-date documentation on third-party code

Read the CLAUDE.md file before you do anything.

$ARGUMENTS

Then in Claude Code:

claude
> /go implement OAuth with Supabase

This gives Claude:

Real-World Impact

One developer's report:

"60-70% time savings on complex features. The combination of Serena understanding my codebase and Context7 providing current docs means Claude writes code that actually works and follows our patterns."

These MCP servers align perfectly with Rendle's four-persona workflow:

Technical Deep-Dives Discussed

Claude Code - Recent Major Updates

Version 2.1.0 Release (January 2026) - Major Update:

MCP Tool Search - "Lazy Loading" (Mid-January 2026):
Revolutionary change to how Claude Code manages tools:

Model Improvements:

Setting System Prompts in Claude Code

Primary Methods:

  1. CLAUDE.md Files (Most Common) - Hierarchical:

    • Global: ~/.claude/CLAUDE.md (applies to all projects)
    • Project: CLAUDE.md in project root (shared with team, checked into git)
    • Local: CLAUDE.local.md (personal, git-ignored)

  2. Command Line Flag: claude --append-system-prompt "instructions"

  3. Settings JSON: In .claude/settings.json or ~/.claude/settings.json

  4. Custom Slash Commands: Create .claude/commands/your-command.md

  5. Hooks: In .claude/settings.json for automated actions

Important Notes:

Best Practices for CLAUDE.md:

OpenSpec vs Claude Code Plans

Claude Code Plans (Built-in):

OpenSpec (External Framework):

OpenSpec Workflow:

  1. Proposal (/openspec:proposal) - Create change proposal
  2. Fast-forward (/openspec:ff) - Generate all planning docs: proposal.md, specs/, design.md, tasks.md
  3. Apply (/openspec:apply) - Implement the tasks
  4. Archive (/openspec:archive) - Move to archive and consolidate into permanent specs

Key Differences:

Claude Code Subagents for Rendle's Four Personas

Implementation: Create four subagents using /agents command or manually create files in .claude/agents/:

Workflow in Action:

> Use the analyst agent to gather requirements for a user authentication system
# Analyst produces REQUIREMENTS.md

> Now use the architect agent to design the solution based on REQUIREMENTS.md
# Architect produces ARCHITECTURE.md and IMPLEMENTATION_PLAN.md

> Use the developer agent to implement the authentication system
# Developer writes the code

> Use the reviewer agent to review the implementation
# Reviewer provides feedback

Advantages Over Rendle's Approach:

Pro Tips:

Conclusion: Claude Code subagents are much better implementation of Rendle's concept. They provide same role separation, better automation, better state management, better team collaboration, better isolation, better security.

Skills vs Slash Commands in Claude Code

Core Difference:

Detailed Comparison:

Aspect Slash Commands Skills
Invocation Manual Automatic
Who Controls User-driven AI-driven
Location .claude/commands/ (single .md file) .claude/skills/ (can be directory with supporting files)
Structure Simple markdown file Can include scripts, templates, reference docs
Discoverability Tab completion in terminal Claude loads descriptions into context
Use Case Repeatable workflows you want to control Background knowledge/procedures Claude should know
Latency Fast (1-2 seconds) Slower (3-5 seconds due to tool calling)
Can Bundle Code No Yes
Cross-platform Claude Code only Works in Claude.ai, Claude Desktop, Claude Code

When to Use Each:

Use Slash Commands When:

Examples: /deploy production, /create-component Button, /run-tests unit, /format-code, /git-commit

Use Skills When:

Examples: API design patterns, security best practices, testing conventions, code style guides, legacy system documentation

The Reliability Problem:
Skills don't always activate reliably. Many developers report:

Consensus: Start with slash commands, use skills for passive knowledge.

Important Update: v2.1.3 Merge (January 2026):
As of Claude Code v2.1.3, slash commands have been merged into the skills system. Old slash commands in .claude/commands/ still work, but now they're part of unified skills system.

New unified format:

Real-World Usage Pattern:
Most experienced developers follow:

  1. CLAUDE.md (project conventions, always active)
  2. Skills (API conventions, security patterns, domain knowledge - Claude decides when to apply)
  3. Slash Commands (/deploy, /create-component, /run-tests - you explicitly invoke)

The Bottom Line:

Contacts

Chris Ayers

Role: Senior Software Engineer, Azure CXP AzRel, Microsoft
Team: Azure Customer Experience Platform - Azure Reliability
Work: Builds AI-driven tooling to analyse outages and incidents across Azure

Contact:

Key Takeaways

Architecture & Resilience

  1. Messaging Patterns are Critical

    • Don't just use pub/sub - understand Outbox, Saga, Scatter-Gather patterns
    • Pattern selection drives operational resilience
    • AWS Step Functions validates Saga orchestration understanding

  2. Failure is Always an Option

    • Design for failure, not just success
    • Use uptime.is to understand real impact of availability targets
    • 99% vs 99.99% is massive engineering effort difference

  3. Azure Resilience Ecosystem

    • Architecture Centre: Reference architectures for what to build
    • APRL: Automated scanning to validate resilience
    • Azure Verified Modules: Pre-built, battle-tested IaC
    • All three work together for comprehensive resilience strategy

AI & Agent Development

  1. Specialized Agent Pipelines > Single Agent

    • Rendle's four-persona approach (Analyst → Architect → Developer → Reviewer)
    • Mimics proven human SDLC patterns
    • Clear separation of concerns, better context management
    • Claude Code subagents are native implementation of this pattern

  2. MCP Servers are Game-Changers

    • Serena: Semantic code understanding (90% of Cursor functionality, free)
    • Context7: Real-time, version-specific documentation
    • Together eliminate: outdated APIs, context overload, poor code quality

  3. AI Agents Need Proper Security

    • ABAC > RBAC for AI agents
    • Consider WHO, WHAT, WHEN, WHERE, WHY
    • "Giving AI agents broad permissions = giving toddler car keys"
    • Build in observability, approval gates, least privilege

Development Practice

  1. Native Web APIs are Here

    • Popover API and Dialog element exist
    • Better accessibility than JS libraries
    • Progressive enhancement opportunity

  2. Usability Testing is Essential

    • Most cost-effective UX research method
    • User testing ≠ usability testing
    • Critical for government digital services

  3. AI Observability Matters

    • Traditional logs insufficient for AI agents
    • OpenLit for LLM/agent workload monitoring
    • Essential for RAG systems, multi-step agents

Incident Management

  1. Incident Plans Need Testing
    • CrowdStrike incident validated 3-year refined plan
    • Post-incident reviews drive improvements
    • Public sector must have robust incident processes

Cross-Framework Knowledge

  1. Cloud Frameworks are Similar

    • AWS and Azure Well-Architected Frameworks share 5 core pillars
    • Principles are cloud-agnostic
    • Understanding one makes learning other straightforward

  2. Code is Human Expression

    • "burn baby burn" in Apollo 11 code
    • Developers leave traces of themselves in code
    • Software is human, not just functional

Tools & Productivity

  1. Claude Code Evolution

    • MCP Tool Search revolutionized context management (134k → 5k tokens)
    • Opus 4.5 inflection point for autonomous coding
    • Skills vs slash commands: Use both strategically

  2. Infrastructure as Code Maturity

    • Azure Verified Modules set standard
    • Officially supported, regularly updated
    • Evolution from CARML/TFVM to unified AVM

Government/Public Sector Relevance

High Priority Applications:

Strategic Value:

Next Steps

Immediate Actions:

  1. Install Serena and Context7 MCP servers in Claude Code
  2. Test Rendle's four-persona workflow using Claude Code subagents