AI agents fail when small decision, integration, heuristic, and preference errors compound across many steps.
2
Production agents need security evaluations, identity-aware enforcement, observability, and human approval for high-impact actions.
3
Enterprise AI gains value from domain context, good evaluation data, workflow integration, and systems designed for production rather than demos.
Summary
This leadership-day compilation covers the practical work of putting AI into companies. Grace Isford explains why agents still fail on ordinary tasks such as booking a flight, then proposes better data curation, evaluations, scaffolding, user experience, and multimodal interfaces. Other talks cover the organizational and technical barriers around production deployment. Speakers discuss agent security, identity propagation, approvals, monitoring, knowledge graphs, domain-specific model evaluation, and the infrastructure needed for GPU clusters. OpenAI and Anthropic describe how they work with enterprise teams, while Datadog and Arize focus on agent traces, task evaluation, convergence, and operational debugging. Booking.com shares measured gains from AI coding tools, and LinkedIn describes a platform built around orchestration, prompt versioning, skills, memory, and telemetry. The day closes with advice on hiring AI teams and turning RAG into useful production systems through specialization, context, attribution, and ambitious business goals.
Agent reliability breaks down through cumulative errors
Grace Isford argues that agents are still unreliable because small mistakes compound over multiple steps. In her flight-booking example, OpenAI Operator struggled to respect airline, price, traffic, arrival-time, and seating requirements. She separates the failures into decision errors, implementation errors, heuristic errors, taste errors, and the mismatch between human expectations and agent speed. A system that is 95% accurate at each step can diverge sharply from one that is 99% accurate after 50 consecutive tasks. Her response is to curate proprietary and generated data, collect personal preference signals, build evaluations for both verifiable and subjective tasks, add scaffolding and checkpoints, and design interfaces around the user's workflow.
Bad AI strategy starts with organizational division and vague promises
Hamel Husain and Greg Ceccarelli present failure advice by inversion. Their examples include promising customers that AI will do everything, buying expensive GPUs without cost analysis, creating tangled systems that only executives understand, and declaring an unclear ambition such as becoming the global AI leader. They warn against hiding work behind jargon, excluding domain experts by calling every task an agent problem, shipping untested bots directly to production, and treating every problem as a reason to buy another tool. They also criticize dashboards filled with generic metrics such as cosine similarity, BLEU, and ROUGE when those measures do not track business outcomes. Their deliberately sarcastic guide points toward clearer strategy, user involvement, relevant tests, and direct inspection of data.
Agent security requires risk evaluation, enforcement, and observability
Don Bosco Durai describes agents as autonomous systems that reason, choose workflows, call tools, access databases, and share memory. Many frameworks run agents, tasks, and tools in one process, which can expose service credentials and tokens across components. The unknown workflow chosen by an autonomous agent also creates a wider attack surface than traditional software. He proposes three layers. Pre-production evaluations should test prompt injection, data leakage, unauthorized actions, vulnerable dependencies, and runaway loops, then produce a risk score. Enforcement should propagate user identity from the initial request through every task and tool to the final data or API access, with authentication, authorization, sandboxing, guardrails, and approval thresholds. Observability should track failures, sensitive data movement, changing models, user behavior, and anomalies in near real time.
Coding agents become useful when they have context and a safe execution loop
Colin Flaherty says more than 90% of Augment Code's roughly 20,000 lines of agent code were written by the agent with human supervision. It added integrations, wrote tests, profiled its own slow code, and used a Google search integration to learn the Linear API before implementing a Linear integration. The system combined codebase retrieval, file editing, subprocess management, clarification, and memory. Flaherty cautions against treating agents as senior engineers or dividing them into narrow categories such as backend and frontend agents. Their strengths differ from human strengths, and capability is better understood as levels of task complexity. A knowledge base helped the agent learn tools such as Graphite, internal test commands, and style rules. He also found that giving the agent the ability to run tests produced a much larger bug-fixing improvement than a foundation-model upgrade alone.
Enterprise RAG works better when graphs preserve relationships and provenance
Stephen Chin and Jonathan Lowe describe Pfizer's technology-transfer use case, where teams must move knowledge from laboratory work into factories. They loaded document chunks into a knowledge graph and used the structure to improve how documents were chunked and retrieved. Lowe says graph-based organization also made it faster for engineers and data scientists to understand a new data landscape. The architecture combines vector search with graph traversal, bringing semantically similar material together with related entities and relationships. The speakers argue that this can reduce generic answers and hallucinations, provide more precise context, support access controls on graph properties, and make answers easier to explain. The business case matters because faster transfer of manufacturing knowledge can help life-science teams move products from research into production.
AI coding ROI needs product metrics and developer education
Booking.com's Bruno Passos and Sourcegraph's Beyang Liu describe a coding-assistant rollout across more than 3,000 developers. Booking.com's data-heavy culture had left feature flags and dead code in a large codebase, creating maintenance toil. The team moved beyond the weak metric of self-reported hours saved and defined measures around lead time, code quality, codebase insights, and modernization. By November, daily Cody users were shipping 30% more merge requests, and the team began building agents for GraphQL generation, large-scale code migration, and organization-specific code review. A joint hackathon helped turn vague experiments into working systems. The speakers stress that education was a major factor: developers who initially stopped using the tool often became daily users after workshops and hands-on help. They also propose declarative service rules that agents can enforce in review and inside the editor.
Agent evaluation must cover routing, skills, paths, memory, and audio
Aparna Dhinkaran breaks an agent into a router, skills, and memory. The router decides which skill to call, a skill executes a chain of model and API operations, and memory preserves state across turns. Each layer can fail independently. Teams should test whether the router selects the correct skill and parameters, whether retrieval chunks and generated answers are relevant and correct, and whether a skill converges on a reliable number of steps rather than wandering through long trajectories. Voice agents add evaluation for speech-to-text accuracy, audio quality, intent, and consistent tone. Dhinkaran shows traces from Arize's own copilot, with evaluations at several points in the flow rather than only on the final answer. She recommends repeated tests with the same skill and varied versions of an input to measure convergence and expose failures.
Production agent teams need narrow tasks, living evaluations, and readable observability
Diamond Bishop describes Datadog's AI on-call engineer and AI software engineer. The on-call agent reads runbooks, logs, metrics, and traces, forms hypotheses, tests them with tools, suggests remediation, and can prepare a postmortem. The software engineer investigates errors, proposes code fixes, writes tests, and can open a pull request or a VS Code diff. Bishop's main lesson is to define a job to be done at a task level that can be measured and verified. Datadog favors vertical task-specific agents over a generalized agent. Evaluations should be designed before the demo, run offline and online, and become a living test set. Observability needs an agent-oriented view because a single workflow can contain many model calls, tool calls, loops, and decisions. An agent graph can make the failing node easier for humans to inspect.
Enterprise AI adoption depends on context, workflow fit, and ambitious use cases
Douwe Kiela frames enterprise AI around a context paradox. Language models can write code and solve difficult problems, but they struggle to place information in the right organizational context. He argues that the model is often only 20% of a larger RAG system, so teams should optimize the full system rather than chase every new model. Company expertise and noisy internal data are sources of differentiation, while pilots are easy compared with production across large document sets, many users, security requirements, and many use cases. Kiela recommends designing for production from the start, giving barely functional systems to real users early, integrating them into existing workflows, and making the first useful moment easy to reach. Accuracy is only a baseline. Attribution, audit trails, claim checking, and observability help handle the remaining errors. He urges teams to aim at business problems with real ROI instead of small internal conveniences.
"The goal isn't really to build a complicated system, it's just to build a system that works."3:14:19
Who should watch
You are responsible for moving agents from demos into production and need concrete checks for security, evaluation, monitoring, and user adoption.
Your company has a large, messy codebase or internal knowledge base, and you need to decide where coding agents, graph RAG, or workflow automation can produce measurable value.
You are building an AI platform or hiring an AI team and want guidance on orchestration, memory, skills, developer experience, infrastructure, and recruiting.
