Task tracking

A structured task list, held outside the agent's context window, is the plan's external source of truth

A task tracker holds that list where both sides can see it: the agent reads and writes it between turns, and so do you, as work proceeds. It records what’s done, what’s in progress, and what’s still to come.

Without one, the plan exists only in the conversation. As the conversation grows, earlier sub-tasks may leave the context window, and “done” can mean only that the agent claimed completion. A tracker records the complete task list and lets both you and the agent compare each progress report with the file.

Markdown and file-based trackers#

• MrLesk/Backlog.md — markdown task tracker with an MCP server, kanban view, and AI-agent guidelines. Tasks live as markdown files in backlog/, easy to read, grep, and version with the rest of the repo.
• steveyegge/beads — Steve Yegge’s “memory upgrade for your coding agent.” A distributed graph issue tracker (SQLite + Dolt-backed) with bd ready for dependency-aware next-task selection, hash-based IDs to avoid merge collisions, and semantic compaction of closed tasks. JSON output for agent consumption.
• hmans/beans — Hendrik Mans’s flat-file alternative inspired by beads: plain markdown files in .beans/, CLI-driven, designed so the agent runs beans prime at session start to rehydrate the plan. Lighter and more human-readable than beads.

Spec-driven workflows produce task lists too#

The same pattern shows up inside the spec-driven toolkits — spec-kit, superpowers, cc-sdd, agent-os; see the catalogue’s spec-driven section — which take a brief or PRD and emit a tracked task list as the Tasks stage of a longer pipeline, rather than offering the tracker as a standalone tool.

MCP-server task lists#

• eyaltoledano/claude-task-master — “Taskmaster”: npm-installable MCP server that drops into Cursor, Cline, Windsurf, and Claude Code. Parses a PRD into tasks; exposes next_task, expand_task, and dependency tracking. Probably the most-installed tool in this category.
• floriscornel/todo-mcp — generic project-scoped task-list MCP server with priorities and persistent storage. Minimal example of “give the agent a todo list it can call as a tool.”

Pick one and keep it boring#

Use whichever tool you will update consistently. For coursework, keep one tracker per prototype across sessions. On the morning of your cutoff, it should show exactly what remains before the crit.

Use separate tools for separate records. The agent’s tracker (beans, beads, Backlog.md) is working memory that can be rewritten freely. A human-visible tracker (Linear, GitHub Issues, or a course assignment brief) is the permanent record across sessions.