Talk Wall

Developing with AI agents like Claude, Codex, or Cursor can quickly turn into a troubleshooting loop. This talk shows how Backlog.md brings structure back to AI-assisted development with smaller, reviewable tasks and a repeatable spec-driven delivery flow.

We walk through the full loop in practice: define tasks, execute with agents, and review outcomes with confidence. The goal is to make AI-assisted feature delivery more deterministic, predictable, and easier to scale.

The landscape of AI coding tools is exploding. We evolved quickly from simple autocomplete to the GitHub Copilot era, and now we face a new wave of Agentic CLI tools, protocols like MCP, ACP and A2A, and frameworks like ADK. It moves so fast that catching up feels impossible.

This talk aims to ground us. I will guide you through this explosion of tools, separating the hype from the reality. We will focus on practical applications and identify the specific tools that actually help us ship real software right now.

Never leave your terminal to create and manage tasks for your AI agents. Backlog.md stores all your tasks as Markdown files in your Git repo. By exposing the main workflows and commands as MCP tools, your AI agents will know how to take tasks from “To Do” to “Done,” and you will no longer run out of context window or miss important requirements in any of your features.

This talk shares the journey from early AI coding struggles to a reliable workflow using Backlog.md. We explore why Markdown tasks with clear acceptance criteria are the missing link for effective agent collaboration, enabling a 3-way review process (Spec, Plan, Code) that keeps agents on track.

In this workshop, we explore the ‘Spec-driven AI Development’ methodology. We’ll see how defining tasks in Markdown with clear Acceptance Criteria and Implementation Plans allows AI agents to work reliably. Participants will get hands-on experience initializing Backlog.md, creating tasks, and guiding agents through the 3-way review process (Spec, Plan, Code) to achieve a high success rate.

This talk traces the journey from initial struggles with AI coding assistants to mastering effective workflows using Backlog.md. It shares best practices for organizing tasks, maximizing AI success rates, and safely integrating code contributions, giving developers repeatable checklists for agent-driven delivery.

Backlog.md is a lightweight, Git-native CLI tool for managing project tasks offline using Markdown files. Designed for solo developers and small teams, it enables seamless collaboration between humans and AI agents without extra infrastructure. Features include a terminal Kanban board, offline web UI, and cross-platform support.

This is the same loop that built Backlog.md (99% agent-written) and pushed it past 3k GitHub stars.

We move from vibe coding slop to a simple spec-driven loop and see why Backlog.md (tasks as markdown in your repo) gives agents the right understanding and context. No prior agent experience needed.