Agentic AI Foundation Video Summaries

A presentation by Juan, a security engineer at Woven (Toyota), on building an AI agent that answers questions and supports decisions on exception requests for internal security standards. The session covers the manual process that preceded the agent, the agent's architecture built with Go and MCP, and the quality evaluation approach used to make its outputs reviewable by humans.

■ Problem and existing workflow
- Roughly 300 internal standards, processes, policies and guidelines that teams must account for
- The manual exception request flow, its security review step, and the friction from multiple time zones

■ Agent design and implementation
- Where the AI agent was inserted and what outputs it produces (approve, reject, request more information)
- Options weighed for context handling: full context, RAG approaches, and fine-tuning
- Context delivery via MCP tools over Markdown standards stored in GitHub, implemented in Go

■ Evaluation and quality control
- Deterministic controls: tool-call verification, output format checks, and a retry loop
- Internal metrics (coherence, specificity) plus Ragas metrics (relevance, faithfulness, precision)
- Processing results across a few hundred tickets and identified areas for further work

■ Q&A
- Retrieval cadence through CI, token usage per ticket, and handling of context size

Engineers and teams building LLM agents, or working in compliance-heavy organizations, can take away a concrete pattern for combining automated assistance with human accountability, along with a practical framework for measuring agent output quality.

Agents spend most of their time waiting, not working. This talk examines how interactive agents that suspend and resume can eliminate idle compute costs, and how the Model Context Protocol (MCP) supports that pattern.

■ Why Agents Idle
- Where waiting happens in a typical agent loop, and the cost of keeping processes running for weeks or months
- Three defining traits of interactive agents

■ MCP Features for Sleepable Agents
- Idle suspension: replacing long-lived client-server streams with multi-round-trip requests
- Work continuing during suspension: the tasks extension and task IDs for long-running calls
- External wake-ups: server-to-client events via polling, data streams, and webhooks

■ Practical Experience
- Amazon's Infinity experiment, which runs agents as short execution slices
- Scaling examples and three calls to action for adopters

Developers building MCP clients and servers, agent framework authors, and teams running always-on agent workloads will come away with a concrete view of the suspension pattern and the spec updates that enable it.

As AI agents increasingly act as end users, product teams need to decide how those agents can actually complete tasks on a user's behalf, not just discover features. This talk introduces Agent Experience (AX) and compares the three main doors agents can use to execute tasks — MCP servers, WebMCP, and existing command-line interfaces — covering how each one works in live demos.

■ Agent Experience (AX) and Why It Matters
- The gap between an agent finding a feature and being able to authenticate and complete a task for the user
- Traffic data cited from Vercel and Cloudflare on the growth of AI bot traffic relative to human web traffic
■ The Three Doors: How Agents Execute Tasks
- MCP: agent-facing interface layer with structured tools, local or remote servers, and a unified OAuth 2.1 authorization path
- WebMCP: lifting a site's existing HTML/JavaScript capabilities into a programmatic tool layer, demonstrated with ChatGPT interacting with a website
- CLI: reusing already-installed command-line tools through an agent, demonstrated with deploying an application via terminal commands
■ Comparing the Doors
- Capability leverage, user-visible feedback, access control and authentication, distribution channels, and composability across the three approaches
- Guidance that the choice of door is an AX decision based on your audience's capabilities, preferred distribution, and integration needs
- Quick design tips for MCP servers and CLI interfaces, plus a mention of WebMCP's experimental status and upcoming release

This session is aimed at developers, product teams, and engineers building agent-accessible products. Viewers will come away with a framework for choosing between MCP, WebMCP, and CLI, along with practical design tips for exposing their own services to agents.

حلقة تناقش أحدث تطورات بروتوكول MCP في مؤتمر Agent Con و MCP Con، بالإضافة إلى قصة BVNK (شركة مملوكة لـ Mastercard متخصصة في العملات المستقرة) وكيفية تبني الذكاء الاصطناعي و MCP داخل المؤسسة بفريق هندسي صغير.

■ أخبار MCP و مؤتمر Agent Con
- ملخص حدثي Agent Con و MCP Con في سان خوسيه والمقارنة بينه وبين حدث أمستردام
- المشاريع الأربعة الجديدة: Agent Skills و Open Responses و Open ENV و Open Sandbox
- حالة مواصفات MCP القادمة وجدول مؤتمرات MCP Developers Summit للعام القادم

■ تجربة BVNK في تبني MCP
- نبذة عن BVNK و حجم الفريق والاستحواذ عليها من Mastercard
- الأدوات المستخدمة في البداية ونقطة التحول مع MCP
- إنشاء بوابة MCP مركزية وإدارة الأذونات والمراقبة
- أثر MCP على فرق غير تقنية مثل المالية و الامتثال و التذاكر
- نصائح عملية لمن يفكر في تطبيق MCP كوظيفة مركزية

هذه الحلقة مفيدة لمهندسي البرمجيات وقادة التقنية وأصحاب المنشآت الراغبين في فهم كيفية تطبيق MCP بشكل عملي داخل المؤسسات، مع أمثلة واقعية من بيئة عمل فعلي.

Shawn, who leads MCP at Hugging Face and co-chairs the transport working group, walks through the evolution of MCP transports—from early local setups to Streamable HTTP and the stateless July release. The talk covers the practical challenges of remote MCP in production, how the community collaborated on protocol improvements, and what the new release means for SDK maintainers and developers.

■ Early MCP and the Shift to Remote Transport
- Early workflow: installing servers from NPM or GitHub, editing JSON config, and connecting locally
- LSP-inspired one-to-one client/server assumptions and why the initial remote transport was impractical
- March release of Streamable HTTP and the user-experience impact of a single URL
- Production challenges: session management, elicitation timeouts, and protocol overhead (73 messages per tool call in June data)

■ The Stateless July Release
- Community collaboration involving Anthropic, Microsoft, and Google contributors
- Removal of lifecycle handshakes and initialization calls, plus built-in list caching
- New call design pattern, HTTP-level routing visibility, and deprecation policy
- SDK maintenance effort that enabled over 50% of tool calls to move to the new protocol

■ Adoption and Deployment Guidance
- Migration status across Anthropic, OpenAI, Hugging Face chat products, and other clients
- SDK version upgrades, enabling list caching, and reconsidering discovery endpoints
- Optimistic connection handling versus pessimistic error handling
- Using MCP to move computation and storage access to where large datasets reside, aided by HTTP routing

Recommended for developers and teams working with MCP servers or agent tools. It helps you understand what changed in the protocol, assess your own migration timeline, and plan concrete steps for upgrading SDKs and adopting new transport features.

A compilation of keynote talks from AgentCon North America covering how agentic AI systems are being built, governed, and connected through open protocols such as MCP. Speakers from industry and research discuss protocol security, agent-to-agent collaboration, enterprise deployment, software correctness, and emerging economic mechanisms.

■ Protocol and platform updates
- MCP specification changes, tool-list notifications, and authorization considerations
- Enterprise agent governance, identity, registries, observability, and gateways
- MCP adoption across Claude products and directory listings

■ Agent collaboration and software development
- Personal agent networks and cross-owner agent rooms
- Agent-written code review, session logs, and Git infrastructure for agents
- Universal properties and formal methods for correctness in agent-generated software

■ Applied examples and ecosystem
- A computationally designed mRNA cancer vaccine for a dog
- PyTorch continual learning loops and on-the-job training for models
- Mechanism design and open-source lessons for the AI economy
- Community awards and foundation updates

The recording is a conference keynote compilation rather than a hands-on tutorial. It is suited to developers, platform engineers, AI product leaders, and researchers interested in agent infrastructure and open standards. Viewers gain an overview of current agentic AI challenges and approaches, and can use the talks to assess protocol security, governance, and collaboration models for their own projects.

Daniel Oh, a developer at Red Hat and IBM, presents an approach to building sovereign, open-source AI coding agents rather than relying solely on commercial cloud AI services. The session covers the Agent Tech AI foundation ecosystem, local open models and runtimes, security guardrails, and a live OpenCode demo that fixes an input-validation flaw in a Quarkus Java application.

■ Open-source agent ecosystem and local models
- Agent Tech AI projects including MCP, Agent Gateway, Goose, AGENTS.md, and A2A
- Open models such as Qwen 3/Qwen Coder, Granite, Llama, and Mistral
- Local runtimes and deployment options including Ollama, LM Studio, Podman, Container AI Lab, and Kubernetes

■ Secure coding-agent workflow
- CLI assistants such as OpenCode, Cloud Code, OpenAI Codex, Google Anti-Gravity, Microsoft Copilot, Continue, and Cline
- Human-in-the-loop review before commit/push and CICD/GitOps deployment
- Guardrails for credential redaction, command filtering, AST analysis, and prompt/SQL injection risks
- Trade-offs across data sovereignty, latency, and cost for local versus cloud approaches

■ Live demo
- OpenCode configured with a Qwen 27B model on Kubernetes
- A Quarkus Java game app with an input-validation issue
- AI-assisted code changes and test updates, with manual approval before push

Daniel Oh, Red Hat and IBMの開発者であるDaniel Oh氏は、商用クラウドAIサービスだけに依存するのではなく、主権的でオープンソースのAIコーディングエージェントを構築する方法を紹介します。セッションでは、Agent Tech AI foundationのエコシステム、ローカルのオープンモデルとランタイム、セキュリティガードレール、そしてQuarkus Javaアプリケーションの入力検証の欠陥を修正するライブOpenCodeデモを取り上げます。

■ オープンソースのエージェントエコシステムとローカルモデル
- MCP、Agent Gateway、Goose、AGENTS.md、A2AなどのAgent Tech AIプロジェクト
- Qwen 3/Qwen Coder、Granite、Llama、Mistralなどのオープンモデル
- Ollama、LM Studio、Podman、Container AI Lab、Kubernetesなどのローカルランタイムとデプロイオプション

■ 安全なコーディングエージェントのワークフロー
- OpenCode、Cloud Code、OpenAI Codex、Google Anti-Gravity、Microsoft Copilot、Continue、ClineなどのCLIアシスタント
- コミット/プッシュ前のHuman-in-the-loopレビューとCICD/GitOpsデプロイ
- 認証情報の秘匿化、コマンドフィルタリング、AST分析、プロンプト/SQLインジェクションのリスクに対するガードレール
- ローカルとクラウドのアプローチにおけるデータ主権、レイテンシ、コストのトレードオフ

■ ライブデモ
- Kubernetes上でQwen 27Bモデルを設定したOpenCode
- 入力検証の問題があるQuarkus Javaゲームアプリ
- AI支援によるコード変更とテスト更新、プッシュ前の手動承認

このセッションは、プライベートまたはローカルでホストされるAIコーディングアシスタントを評価している開発者、プラットフォームエンジニア、セキュリティまたはコンプライアンスの関係者を対象としています。視聴者は、エージェント型AIツールを採用する際に検討すべきアーキテクチャ、セキュリティ管理、運用上のトレードオフについて構造化された見方を得られ、さらにクローンしてテストできるデモリポジトリも得られます。

AI agent adoption often outpaces the ability to prove business value, and this talk examines that gap through a practical evaluation lens. It follows an open-source evaluation framework called Higher and a real incident-response agent example to show how teams can decide whether to hire, train, or retire an agent before scaling it. The discussion focuses on evaluation and product decision-making rather than model architecture or vendor comparisons.

■ Human readiness and organizational change
- Human and habits: whether teams, tools, and data are ready to trust an agent
- Why, Impact, Intent: defining an agent's job description, scope, and ROI

■ Choosing the right solution
- Rules vs responses vs reasoning: matching task ambiguity to the appropriate automation level
- Guardrails and MCP boundaries: human-in-the-loop controls for high-risk actions

■ Evaluation metrics for probabilistic software
- Quantitative metrics such as edit distance, ping-pong rate, do-it-myself rate, instant resolution rate, and red-line rate
- Qualitative triage using deterministic telemetry, LLM classification, and unsupervised topic modeling
- Tone and human feedback as signals for product fixes and adoption

■ Practical implementation
- An incident-response SRE agent case and a GitHub repository with guidance, metric definitions, and a sample analytics workflow

Viewers responsible for AI product strategy, analytics, or engineering management gain a structured way to assess agent value, safety, and adoption. The talk is suited to teams planning or already deploying AI agents and offers an action plan: test the agent idea against the Higher framework, define measurable outcomes, and use the shared repository to prototype evaluation before production.

يناقش المتحدث تطور نماذج الذكاء الاصطناعي خلال العامين الماضيين، من نماذج تكتفي بالدردشة إلى وكلاء يتخذون إجراءات فعلية، موضحاً كيف يتجلى هذا التحول في مكونات الوكيل: الاستنتاج، واستدعاء الأدوات، والتخطيط. ويبيّن الانتقال من التلقين اليدوي عبر الموجهات إلى السلوك المدرب داخل النموذج نفسه.

■ نشأة الوكيل الذكي
- مكونات الوكيل الأساسية: النموذج اللغوي، الاستنتاج، الذاكرة، الأدوات، والتخطيط
- الانتقال من موجهات الاستنتاج اليدوية إلى دمج الاستنتاج داخل النموذج
■ استدعاء الأدوات والتخطيط
- توليد استدعاءات الأدوات بتلقين خارجي مقابل تنفيذها بشكل أصيل
- تشغيل الوكيل في حلقة تكرارية بنموذج الاستنتاج والفعل
- معالجة مخرجات الإخراج يدوياً مقابل التدريب على بيانات التفاعل
■ أثر تفويض المهام على تطوير النماذج
- تطوير النماذج من منظور تفويض المهام البشرية للاستنتاج والأدوات والتخطيط
- التوجهات المستقبلية مثل الأسراب وبيئات العمل الوكيلية وضوابط الأمان والتقييمات

يستهدف مهندسي التعلم الآلي والمطورين ومهتمي أبحاث الذكاء الاصطناعي الراغبين في فهم بنيان الوكلاء وتطوّرها. يتيح الاطلاع على المقارنة بين الأجيال السابقة والحالية من النماذج فهمًا واضحًا لمفهوم الوكيل وما الذي يتغير في تدريبه، مع إمكانية متابعة الكتب المرجعية المذكورة للتوسع في الموضوع.

This episode of The Context features Darren, chief architect at Obot AI, presenting DiscoBox, an open-source tool for running AI coding agents in isolated local or remote sandboxes. The session covers the tool's design goals, security model, and a live demonstration of creating and managing sandboxed environments. It also touches on upcoming AgentCon and MCPCon events and practical installation steps.

■ DiscoBox overview
- Provides a full micro VM or container for AI agents
- Supports local and remote operation with Git-based workflows
- Installs via curl or brew on macOS, Windows, Linux, and WSL

■ Security and credential management
- Uses sentinel tokens instead of real secrets
- Scopes credentials to specific use cases and validates them with AI
- Allows time-limited access and auto-approval mechanisms

■ Live demo highlights
- Creating sandboxes, cloning repositories, and running full applications
- Accessing shell and a desktop environment (XFCE), plus IDEs like VS Code and Zed
- Running multiple sandboxes in parallel and modifying site colors
- Demonstrating a 'develop' skill that automates tests, reviews, QA, PR, and CI

■ Additional topics
- Upcoming AgentCon and MCPCon on October 22 in San Jose
- Technical challenges during the live stream and platform comparisons

The video is aimed at developers, DevOps engineers, and AI practitioners interested in secure agent sandboxing. It provides a practical overview of DiscoBox's installation, configuration, and credential-handling approach. Viewers can learn how to set up and experiment with isolated environments for AI agents.

Dylan Ratcliffe, founder and CEO of Overmind, explains how his team removed human code review from most production changes and replaced it with reviewed execution plans plus AI-driven loops. He describes the bottlenecks that led to this decision, the concerns about team knowledge and compliance, and the internal experiments that reshaped engineering roles.

The talk covers:
■ Why code review was eliminated
- What peer review previously provided: bug detection, SOC 2 / ISO 27001 compliance, onboarding, and shared system understanding
- Why AI-generated code is difficult to review like human-authored code, and how AI responds to review comments

■ Redefining engineering work
- Moving from detailed specs to high-quality code toward handling ambiguity, constraints, and tradeoffs
- A one-page spec and 24-hour demo rule; engineers resolve ambiguity themselves rather than asking product

■ Plan-first until loops
- Human-reviewed execution plans before automated code generation
- Writer/reviewer AI loop iterates until implementation matches the approved plan
- Peer review shifts before the loop; plans are easier and cheaper to reject than pull requests

■ Implementation, metrics, and tradeoffs
- Internal MCP-based tool with no UI for plan sharing and review
- Three-week comparison: more issues completed, more deploys, shorter lead time, and less context switching
- Rejecting a plan adds about 40 minutes; rejecting a pull request adds about 16 hours
- Compliance through pre-approved changes and sampling instead of line-by-line review
- “How” decisions still matter, and engineers should keep making them
- About 8% of planned work is abandoned after planning, which the speaker presents as a positive option
- The internal add-on is published at until.dev

The talk is aimed at engineering leaders, senior engineers, and product managers exploring AI-assisted development and review workflows. Viewers gain a framework for moving peer review earlier, clarifying which decisions should remain human, and measuring the impact of loop-based automation. It also points to the published until.dev add-on for those who want to examine the internal approach.

I demonstrate how to use Logfire to inspect agent runs and identify opportunities to improve an AI agent after deployment. A travel-replanning agent provides a concrete example, covering prompt and tool-description changes, runtime errors, and managed variables.

■ Agent monitoring and improvement
- Reviewing runs, costs, traces, tool calls, and runtime exceptions in Logfire
- Applying suggested changes to prompts and tool descriptions through managed variables
■ Related capabilities
- Using the MCP server for agent monitoring tasks such as dashboards and alerts
- Discussing evaluation-case generation as a feature under development

For developers building Python-based AI agents, this walkthrough offers a practical view of post-deployment monitoring and iteration. You can use the approach to review your own agent runs and decide which prompt or runtime changes to make; the presentation is a product walkthrough rather than a full implementation tutorial.