Reading groupAI REWIND 2025 - MLOps Reading Group Year-end SpecialPushed backRohan Prasad challenged the assumption that adding all available information to ever-larger context windows reliably improves LLM results.15:10
Threads / Agents in production / 2026
Agents in production in 2026
85 sessions
Reading groupAI REWIND 2025 - MLOps Reading Group Year-end Special
Building Agentic Tools for ProductionPushed backSam Partee rejected the idea that all tools should be put in one MCP gateway and said agents should not have too many tools.19:26
PodcastEnterprise AI Operations: The Missing PiecePushed backStarting with one agent and scaling sequentially is not the only valid approach; parallel development can work when processes are standardized and repeated across departments.9:42
Inside OpenAI's AI Agent Collaboration SystemClaimEvals are structured, objective ways to measure an agent's or model's performance, while graders evaluate the model's outputs, reasoning, and justification.1:06
Tool definitions are the new Prompt EngineeringPushed backAlex Salazar argued that most people should currently use carefully selected, specific tools, while the long-term goal is to let agents choose from many tools.11:49
MCP Security: What Happens When Your Agents Talk to Everything?ClaimMCP is a universal adapter that lets AI applications connect to external data sources, tools, and services without custom integration code for each one.0:52
Multi-Agent Systems for the Misinformation LifecycleClaimMisinformation is information that is factually incorrect, misleading, or taken out of context regardless of intent.1:05
Structured Dissent Patterns for Agentic Production ReliabilityClaimPhil Stafford says structured dissent uses multiple agents with different perspectives to produce answers that are more nuanced and closer to ground truth.1:16
Context Engineering pitfalls for our e-commerce agentPushed backThe speakers reject the idea that a larger context window means everything should be placed into it, arguing that excess context can harm model performance.23:33
Expanding context engineering to the tooling layerPushed backFrank Wittkampf disputes the idea that MCP itself filters or optimizes the tool surface presented to an agent; he says MCP connects and standardizes services but does not perform that filtering.3:10
Feedback Loops for Agentic WorkflowsPushed backTudor rejects the idea that reinforcement learning by itself will make an agent optimize correctly.13:00
PodcastGraph Neural Networks Just Solved Enterprise AI?ClaimAgents need predictive AI tools because business decisions such as preventing fraud, retaining customers, and approving loans depend on predictions grounded in enterprise data.42:53
What if AI Doesn't Need Structure, It Needs Connection?ClaimFunction calling created a path for AI to be embedded in existing applications and marked the beginning of AI agents.9:57
Co-Engineering: The New Era of Human-AI CollaborationPushed backWhether AI coding agents can replace software engineering is disputed: some users describe programming or software engineering as dead, while experienced engineers say production use is limited.1:38
Coding with AIClaimChip Huyen says AI coding tools commonly have IDE-based, CLI-based, GitHub-based, and web interfaces.1:39
Stop Building AI Like Traditional SoftwarePushed backAishwarya Naresh Reganti said prompting should be a last resort for controlling agent access, preferring deterministic guardrails and rule-based access controls.16:30
PodcastThe Semantic Layer and AI AgentsPushed backWhether semantic logic should be embedded inside a BI tool or maintained independently.4:09
Tool CallingPushed backAlex Salazar disputes the idea that the main solution to MCP problems is changing how servers are called, arguing that the central issue is building better MCP servers.17:42
Beyond the Gold Standard: Evaluating and Trusting Agents in the WildClaimSanjana Sharma says that reliable production agents require more than capable models because they must handle edge cases, ambiguous requirements, conflicting data sources and imperfect user behavior.0:42
Rethinking Notebooks Powered by AIClaimMarimo can add context about a referenced data frame, including its columns, row count, and column types, so an LLM can write better SQL or data-frame code.8:57
Building Artificial Engineering Intelligence for Engineering TeamsClaimAishwarya Shankar says engineering intelligence should cover background agents, code review, and engineering performance reporting.1:21
MCP Security: The Exploit Playbook (And How to Stop Them)ClaimVitor says MCP adoption is accelerating faster than the security protections available for malicious actors.0:27
The Future of Coding: AI Agents & the Next Tech RevolutionClaimRicky Doar disputes the view that AI is much better at greenfield development than brownfield development because newer tools can search existing code semantically and use its patterns.8:49
How AI covered a human's paternity leavePushed backQuinten Rosseel disputes the common emphasis on text-to-SQL benchmarks as the main measure of agent success, arguing that business context is the real challenge.4:36
When Agents Learn to Feel: Multi-Modal Affective Computing in ProductionPushed backAffective modalities should not be limited to text, audio, and images; future agents could use other signals such as smell, taste, and touch.16:59
Agents as Search EngineersClaimAgentic search treats retrieval as part of a stateful control loop involving reasoning strategies, tool calls, and memory rather than as a final endpoint.3:27
Building an Orchestration Layer for Agentic Commerce at LoblawsPushed backOne-to-one translation from OpenAPI APIs to MCP tools did not work well for Loblaw Digital's needs.20:34
Time to become a hackerPushed backMatt Sharp rejects the idea that users should simply trust agents or allow broad permissions, arguing for least privilege and closer monitoring.9:09
Yes, we do need MCPPushed backBen disputes the idea that MCP client retries are generally safe, arguing that clients should retry only when requests are clearly idempotent.25:18
Multi-Agent Personalization with Shared Memory: From Email to Website to ProposalPushed backThe speaker said that adding MCP and function-calling tools does not necessarily solve shared customer understanding because the tools may be called differently by different agents.6:28
Dynamic Contextual Retrieval in Enterprise AnalyticsPushed backDirk Petzoldt argues that the industry should build data companions and end-to-end analytics processes instead of merely adding natural-language interfaces to business-intelligence tools.2:32
Context Engineering 2.0: MCP, Agentic RAG & MemoryPushed backGiving agents direct access to databases through generated SQL is generally a bad idea because of reliability, security, and performance risks.7:13
Enterprise-ready MCPClaimAgents are changing programming from static code written by developers to dynamic programs driven by natural-language instructions.1:27
From Chat Fatigue to Instant ActionClaimThe team learned that agent experiences should move beyond chat interfaces toward instant actions and dynamic interactions.0:28
Open vs Closed Source Agent Infra?Pushed backBen disputed the assumption that open-source tools let teams fix any problem, saying highly abstract tools can quickly become difficult to understand and can create a painful dead end.11:03
Using Agents in Production: Past Present and FutureClaimProsus expects to have exactly 30,000 agents in production across the group by the end of March.0:04
PodcastWrite Reliable Software with TemporalPushed backJohann Schleier-Smith rejected the idea that LLM nondeterminism prevents durable execution, explaining that the agent's control flow is deterministic and LLM results are captured as activity results.24:37
PodcastPerformance Optimization and Software/Hardware Co-design across PyTorch, CUDA, and NVIDIA GPUsClaimChris Fregly says clearing an AI coding assistant's context window can produce large performance improvements.1:04:47
Everything We Got Wrong About Research-Plan-ImplementPushed backDexter Horthy rejected the view that engineers should avoid reading production code and rely only on plans or automated agents.8:50
Lessons from 25 Trillion Tokens: Scaling AI-Assisted Development at KiloClaimKilo is an all-in-one agentic engineering platform.0:27
2026 The Year of Agent OrchestrationClaimZach Lloyd predicts that 2026 will be the year of agent orchestration.0:54
Decomposing the Agent Orchestration System: Lessons LearnedPushed backNiels Bantilan rejects the idea that agents should be expected to be failure-proof and instead argues for cheap failures, fast recovery, and tight feedback loops.26:55
How to Make a Coding Agent a General Purpose AgentPushed backHarrison said the general-purpose agent interface would probably not be the terminal, while discussing more user-friendly interfaces.21:55
Stop Shipping on Vibes: How to Build Real Evals for Coding AgentsPushed backJessica Wang challenged the conclusion that the evaluation proved agentic search was universally better than vector search.22:34
MCP Dev Summit [Day 1]Pushed backDavid Soria Parra rejects the view that MCP is no longer needed and says the protocol is still the integration protocol for enterprise agent systems.29:04
The Coding Agent Multiverse of MadnessPushed backAnkit Mathur disagreed with the idea that enterprises should standardize on one coding tool for everyone, arguing that developers should use the best tool for each use case.5:38
Ship Agents: A Virtual Conference Track 2Pushed backThe agent should not autonomously repair physical manufacturing equipment in the described system; it supplies grounded information so human engineers can perform the repair.21:16
A New Kind of MarketplacePushed backDonné Stevenson is skeptical that agents will routinely hire other specialized digital agents for services.36:20
PodcastThe Modern Software EngineerPushed backDemetrios Brinkmann questioned whether running many parallel agents is as useful as it is portrayed in social-media demonstrations, and Mihail Eric said the idea is often romanticized.27:46
PodcastWhy Agents are Driving Software Development to the CloudPushed backZach Lloyd disputed the idea that moving agents to the cloud mainly means giving them a cloud computer or sandbox.1:35
OpenXData ConferencePushed backVenoth rejected the idea that federated querying would solve agent context access at scale.33:17
PodcastGetting Humans Out of the Way: How to Work with Teams of AgentsPushed backRob Ennals disputes the idea that people should manually QA agent-written features by clicking through the application and reviewing the code.1:53
PodcastThe Creator of Superpowers: Why Real Agentic Engineering Beats Vibe CodingPushed backJesse Vincent rejects the view that agent swarms are generally the right way to organize agentic software development.17:38
Stop AI Agents From SQL Injecting Your DatabaseClaimAveri Kitsch is the technical lead for MCP Toolbox for Databases, a customizable framework for building database tools.1:00
Why AI Agents Shouldn't Replace Your Fraud ModelsPushed backFull agentic decisioning should not directly replace rules engines and machine learning models in high-stakes systems.6:49
Building AI Agents That Survive ProductionClaimHaytham Abuelfutuh says agents should be built to tolerate failures and react to them because failures will happen in production.7:37
PodcastBuilding MCP Before MCP Existed: Inside Despegar's Sofia AgentPushed backDemetrios Brinkmann suggested that the orchestration layer might no longer be needed as models improve, while Nicolas Alejandro Bogliolo said Sofia still needs it for now.38:35
PodcastThe Latency Goldilocks Zone ExplainedClaimILO is a conversational agent that understands a user's identity, preferences, desires, and request to provide suitable food options.1:13
What's Special About Meta's Multi-Agent SystemsClaimA standalone LLM is not enough because the system needs external context, databases, retrieval, and indexing to compare a video with existing content and creators.4:25
Architecting Modern AI SystemsPushed backAlan said that software sandboxes are useful but cannot guarantee that increasingly capable coding agents will not escape them.53:45
Architecting Modern AI Systems: Platforms, Agents, and IntegrationPushed backAllen Roush disputed the idea that software sandboxes can reliably prevent sufficiently capable coding agents from escaping.53:32
PodcastAgents & the $40M Bet on Multiplayer AIClaimStanislas Polu says current agent use is mostly single-player because agents have short task horizons and jagged capabilities.5:19
Context Engineering for Coding AgentsPushed backFausto disputes the idea that a larger context window should simply be filled, arguing that staying focused and restarting or compacting context is usually better.18:30
PodcastAutonomous Agents at Work: From OpenClaw Hype to Enterprise RealityClaimAutonomous agents have a much larger boundary of potential errors because they use tools and act independently.0:35
PodcastVoice Agent Use CasesClaimVoice agents need a balance between preconfigured behavior and developer flexibility because too many controls can make systems difficult to configure.0:30
PodcastLogs Are All You Need: Rethinking Observability with AI AgentsPushed backSherwood Callaway said third-party evaluation tooling was unnecessary for their current work and that their evaluations were home-rolled.39:34
Omnigent: Composition, Control, and Collaboration for AI AgentsPushed backDenny Lee rejects the idea that agents debate merely to decide whether he should pursue a project; he says the purpose is to determine the right way to do it.10:47
Sandboxing, Agent Harnesses, and Agent TeamworkPushed backThe speakers differ over whether people should focus on a main quest or deliberately enjoy agent-enabled side quests.1:17:56
Coding Agents Are Secretly General AgentsPushed backJay Hack disputes the claim that writing a tool would solve the car-wash reasoning failure.45:42
Events Notifications in MCPClaimMCP can currently notify a client only while it holds an open connection, and current notifications usually indicate only that a resource changed rather than describing the change.0:20
JSON Schema 2020-12 and the Contract for ContextClaimOla Hungerford is a principal engineer at Nordstrom and a maintainer of MCP.0:00
Reading groupLoop EngineeringPushed backArthur Coleman disputed the idea that agents should be allowed to run fully autonomously, based on the risk of uncontrolled parallel work and token exhaustion.15:47
MCP Goes StatelessClaimJohn Dellenbaugh initially designed MCP scaling with a load balancer and multiple systems but learned that the older protocol required sticky state and session tracking.0:00
Policy Enforcement and Tamper-Evident Audit ChainsPushed backImran Siddique rejected the idea that all parts of an agent system need to be nondeterministic.17:43
Agentic DX: Bringing IDP into your IDEPushed backAdnan Vahora rejects the idea that a flat agent mesh is sufficient at scale and argues for a master-agent architecture.6:29
Before the Agent Calls: Source-level Findings from 100 MCP ServersClaimMCP servers can be dangerous because they take instructions from language models that can be manipulated more easily than traditional clients.1:08
MCP 2.0: What Changes, What Breaks, and What You Need to Do About ItClaimThe latest MCP release makes the protocol stateless and begins a transition period in which stateful and stateless implementations will coexist.1:35
MCP Release Overview: Stateless and the Big Changes in the New SpecClaimAlex Hancock is a software engineer at Block, a maintainer of the MCP Rust SDK, and a core maintainer of Goose.0:33
MCPs for Observability StacksPushed backDiana said that MCPs and semantic conventions for AI are not yet fully mature and still require validation against real use cases.23:29
Responsible Autonomy: Building Governance Frameworks for AI That Act in the Real World via MCPPushed backThe host interrupts the planned architecture discussion because the session is running out of time and asks questions about agent sandboxes.24:40
Stateless, Yet Durable: MCP Tasks v2ClaimMCP is stateless in V2, while durability remains an important part of MCP Tasks.1:07
The MCP Tasks ExtensionClaimMCP normally uses a synchronous request-and-response model in which a client sends a request and the server returns a result.1:59
What We Learned from Dozens of Enterprise MCP DeploymentsPushed backThe discussion rejects the idea that first-party MCP servers alone are sufficient, arguing that their generic and limited tools may not reflect a company's business.23:32
Reading groupPrompt Injection as Role Confusion: Rethinking Agent SecurityClaimLucas says similar reasoning forgeries can be placed in tool output instead of a user message, causing the model to treat tool content as user or internal reasoning content.6:58