Agentic Workflows

5 videos · Oct 2026

What this trail covers

This post contributes to the understanding of agentic systems by detailing a platform that integrates AI model steps into deterministic engineering workflows. The system uses YAML definitions for workflows, which undergo the same code review process as other code. It specifies tools and access for each step and runs on existing deployment infrastructure. The system separates orchestration, execution, and session state, and includes features like timeouts, rate limits, and locks to prevent issues. This talk specifically highlights the process of enabling non-technical domain experts (financial planners) to define and build these agentic workflows through a structured, iterative process focused on capturing real-world judgment and limitations.

All AI Engineer trails

Oct 2026Agents in Orbit: Workflows, Not Autopilot — Eduardo Simon & Javier Casal, FlywireThis post introduces Orbit, Flywire's internal platform for integrating AI model steps into deterministic engineering workflows. Orbit allows AI-driven tasks to be defined in YAML, subjected to standard code review, and executed on existing infrastructure. Key technical contributions include the separation of orchestration, execution, and session state; implementation of timeouts, rate limits, and locks; per-step restrictions on network access, tools, and skills; and the use of structured outputs to drive workflow branches. The platform supports a range of engineering use cases, with a focus on predictable work staying in code and AI providing judgment, while human review remains integral.
Oct 2026Reliable, Smart Workflows for Finance, Wealth and Insurance — Karthic Rao, PerpetualThis talk introduces a system for building reliable, smart workflows in regulated industries like finance, wealth, and insurance. It addresses the limitations of fixed flowcharts for ambiguous tasks and the reliability issues of purely improvising agents. The proposed design places control flow, evidence, and external actions within a deterministic system, using models for interpretation and judgment at explicit, typed points. A planner translates requests into semantic graphs, and a compiler validates these against firm-defined action contracts, allowing for clarification requests or refusals. Approved plans are immutable, with external actions having separate gates. State machines handle waiting and follow-ups, and uncertain outcomes are routed to manual review. The design targets a ratio of smaller models for most decisions, with frontier models used sparingly, and outlines a roadmap for privately hosted inference and future training.
Oct 2026Just-in-Time Ontologies: What Cheap Intelligence Means for Data — Hrishi Olickel, Southbridge AIThis talk addresses the limitations of universal schemas and confident agents in data transformation for AI. It proposes a 'just-in-time ontology' approach where transformations are built when a specific question is known, preserving context and evidence. The talk uses an Ohio campaign finance entity resolution project as an example, detailing the separation of models for pipeline design, job execution, and system optimization. Key elements discussed include the importance of definitions, held-out cases, disagreement checks, and model priors. It also outlines four requirements for self-optimizing pipelines: stable execution surfaces, bounded budgets, reusable methods, and defined failure modes, advocating for small pilots and independently checked work segments to manage cost, speed, and accuracy trade-offs while maintaining evidence.
Oct 2026Please… Enough with the Sloppy, One-Shot SaaS Replacements — Mahmoud Abdelwahab, RailwayThis talk details the engineering effort behind building an internal application using coding agents to replace existing SaaS tools for marketing content, email, social posts, and launch planning. The core contribution lies in the emphasis on 'product polish' for AI-generated tools, which includes implementing features like background saving, responsive interactions, reversible actions (undo/archiving), and consistent screens. It advocates for a structured codebase with reusable abstractions, packaging recurring design and deployment preferences into agent skills, and establishing a verification loop that exercises the running feature from start to finish, ensuring a usable and desirable end-user experience beyond just passing tests.
Oct 2026We Let Our Financial Planners Write the Agents — Matthew Ragone, RangeThis post details a five-step process for enabling non-technical users (financial planners) to build agentic workflows without writing code. The process involves identifying repeatable, judicable workflows, teaching agents with real cases, packaging judgment into 'skills' with defined limits and questions, rigorous testing, and assigning ownership. It emphasizes capturing the nuances of professional judgment and establishing clear boundaries for agent actions (e.g., drafting but not sending recommendations). The core contribution is a methodology for democratizing agent creation by focusing on workflow definition and judgment packaging, rather than direct coding.
Advertisement