diff --git a/docs/slides/index.md b/docs/slides/index.md index 080cd23a52..5d5d7f9502 100644 --- a/docs/slides/index.md +++ b/docs/slides/index.md @@ -27,7 +27,7 @@ mermaid.initialize({ startOnLoad: true }); | Runs CI/CD pipelines | Orchestrates AI agents | | Debugging code | Debugging agent behavior | -> The role evolves: from coding to orchestrating intelligent agents +> The role evolves: from coding to orchestrating agents --- @@ -40,7 +40,7 @@ mermaid.initialize({ startOnLoad: true }); - **Supervisor** — Monitor running agents and handle exceptions - **Debugger** — Diagnose workflow behavior and improve prompts -> The human stays in the loop while AI handles execution +> Human oversight, AI execution --- @@ -73,17 +73,12 @@ flowchart LR # Continuous Integration to Continuous AI -- **Accessibility review** - Automated WCAG compliance checks - -- **Documentation** - Auto-generate API docs and README files - -- **Code review** - AI-powered PR analysis and suggestions - -- **Test improvement** - Identify missing test coverage - -- **Bundle analysis** - Monitor package size and dependencies - -- **Issue triage** - Automated labeling and prioritization +- **Accessibility review** — Automated WCAG compliance checks +- **Documentation** — Auto-generate API docs and README files +- **Code review** — AI-powered PR analysis and suggestions +- **Test improvement** — Identify missing test coverage +- **Bundle analysis** — Monitor package size and dependencies +- **Issue triage** — Automated labeling and prioritization > @@ -95,8 +90,6 @@ https://github.com/github/gh-aw/issues/1920 # Evolution: LLMs to SWE Agents -## From code completion to autonomous workflows - **2021: GitHub Copilot** - AI-powered code completion **2022: ChatGPT** - Conversational AI assistant @@ -111,7 +104,7 @@ https://github.com/github/gh-aw/issues/1920 # CI/CD with GitHub Actions -YAML workflows as configuration stored in `.github/workflows/` that trigger on events like push, pull requests, issues. +YAML workflows stored in `.github/workflows/`, triggered on events like push, pull requests, or issues. ```yaml on: @@ -129,7 +122,7 @@ jobs: # The "Lethal Trifecta" for AI Agents -AI agents become risky when they combine **three capabilities** at once: +AI agents become dangerous when these **three capabilities** combine: - **Private data access** @@ -145,6 +138,8 @@ AI agents become risky when they combine **three capabilities** at once: ## Safe by design. Useful by default. +> The best developer tools protect you from catastrophe while letting you build something real + --- # From Scratch to MakeCode @@ -157,7 +152,7 @@ AI agents become risky when they combine **three capabilities** at once: These environments share one superpower: -> You're protected from catastrophe — and you can still build something **real and delightful** +> Protected from catastrophe — still building something **real and delightful** --- @@ -232,7 +227,7 @@ Summarize issue and respond in a comment. --- -# Compiled Action Yaml +# Compiled Action YAML ```yaml jobs: @@ -284,7 +279,7 @@ network: tools: web-fetch: --- -Fetch latest TypeScript docs report findings in a comment. +Fetch latest TypeScript docs and report findings in a comment. ``` > Control external access for security @@ -363,8 +358,6 @@ mcp-servers: # Monitoring & Optimization -Track workflow performance and AI agent behavior. - ```sh # View recent runs gh aw logs @@ -404,8 +397,6 @@ Review this PR with context from previous reviews: # Playwright + Upload Assets -## Browser automation for web app testing - ```yaml --- on: @@ -418,11 +409,10 @@ safe-outputs: upload-asset: # Attach screenshots to artifacts --- Test the web application: -1. Navigate to the deployed preview URL +1. Navigate to the preview URL 2. Take screenshots of key pages -3. Check for visual regressions -4. Validate responsive design (mobile, tablet, desktop) -5. Create issue with findings and screenshots +3. Check for visual regressions and responsive design +4. Create issue with findings and screenshots ``` **Use cases:** Visual regression, accessibility audits, E2E validation for SPAs @@ -447,11 +437,6 @@ safe-outputs: # RECOMMENDED: Use sanitized context Analyze this issue content (safely sanitized): "${{ steps.sanitized.outputs.text }}" - -Metadata: -- Issue #${{ github.event.issue.number }} -- Repository: ${{ github.repository }} -- Author: ${{ github.actor }} ``` **Auto-sanitization:** @mentions neutralized, bot triggers blocked, malicious URIs filtered @@ -462,14 +447,10 @@ Metadata: ## Multi-layered defense in depth -GitHub Agentic Workflows implements a comprehensive security architecture with multiple isolation layers to protect against threats. - -**Key Security Principles:** - - Container isolation for all components - Network firewall controls at every layer - Minimal permissions by default -- Separation of concerns +- Separation of concerns (agent, tools, outputs) --- @@ -517,112 +498,48 @@ flowchart TB # Security Layer 1: Coding Agent Sandbox -## Isolated agent process - -**Coding Agent Sandbox:** - -- Agent process runs in isolated container -- Read-only permissions by default -- No direct write access to repository -- Limited system access +**Agent sandbox** — isolated container, read-only by default, limited system access -**Primary Proxy/Firewall:** - -- Filters outbound traffic from agent -- Controls access to MCP Gateway -- Enforces network allowlists +**Primary proxy/firewall** — filters outbound traffic, controls MCP Gateway access, enforces network allowlists --- # Security Layer 2: MCP Gateway -## Central routing with access controls - -**MCP Gateway (mcpg):** - -- Central routing component -- Manages communication between agents and services -- Validates tool invocations -- Enforces permission boundaries - -**Benefits:** +**MCP Gateway (mcpg)** — central routing between agents and services -- Single point of control -- Auditable tool access -- Prevents direct agent-to-service communication +- Validates tool invocations and enforces permission boundaries +- Single point of control — no direct agent-to-service access +- Full audit trail for tool calls --- # Security Layer 3: Tool Sandboxes -## Isolated MCP servers and skills +**MCP servers & skills** — each runs in its own container, non-root, dropped capabilities -**MCP Server & Skill Sandboxes:** - -- Each MCP server runs in own container -- Each skill runs in separate sandbox -- Non-root user IDs -- Dropped capabilities - -**Secondary Proxy/Firewalls:** - -- Additional proxy layers for egress traffic -- Domain-specific allowlists -- Defense against data exfiltration +**Secondary proxies** — egress filtering, domain allowlists, defense against data exfiltration --- # Security Layer 4: Service Access -## Controlled external communication - -**Service Layer:** - -- External services accessed through proxies -- Multiple security controls before reaching services -- Comprehensive audit trail -- Network traffic monitoring +External services accessed only through proxies — multiple controls before reaching any service -**Defense in Depth:** -Even if one layer is compromised, multiple additional security controls remain in place. +> Defense in depth: if one layer is compromised, additional controls remain in place --- # Security Features Summary -**Container Isolation:** - -- GitHub Actions Jobs in VMs -- Separate sandboxes for agent, MCP servers, skills - -**Network Controls:** - -- Proxy/firewall at every layer -- Domain allowlisting -- Ecosystem-based controls (node, python, containers) - -**Permissions:** - -- Read-only by default -- Safe outputs for write operations -- Explicit permission grants - -**Supply Chain Integrity:** - -- Classic: GitHub Actions pinned to commit SHAs at compile time -- Agentic: Protected files guard CI/CD configuration from tampering - -**GitHub Guard:** - -- `min-integrity` restricts agent access by content trust level -- Secrecy & integrity metadata on safe output messages -- Automatic `min-integrity: approved` for public repositories - -**Monitoring:** - -- Threat detection -- Audit logs -- Workflow run analysis +| Layer | Protection | +|---|---| +| **Containers** | VMs + sandboxes for agent, MCP servers, skills | +| **Network** | Proxy/firewall at every layer, domain allowlisting | +| **Permissions** | Read-only default, safe outputs for writes | +| **Supply Chain** | Pinned action SHAs, protected CI/CD files | +| **GitHub Guard** | `min-integrity`, secrecy & integrity metadata | +| **Monitoring** | Threat detection, audit logs, run analysis | --- @@ -662,7 +579,7 @@ Analyze issue and create implementation PR **Visit:** -Security is foundational to GitHub Agentic Workflows. We continuously evolve our security controls and welcome community feedback. +--- # Getting Started (Agentically)