Hypothesis

Most of us have fallen into the productivity trap at some point. It starts innocently: list every task, check a few off, schedule a few more for the future. But priorities rarely stay stable across the actual timeline of work. Updating the task app becomes a task of its own. Soon the overdue list grows, the app starts to feel accusatory, and opening it creates more dread than clarity. Eventually, we start doing work without capturing it - and the whole system drifts away from reality.

I have gone through this loop a number of times. Each attempt starts with optimistic enthusiasm and ends in quiet frustration. The recurring failure modes are:

• Task priorities change faster than the task list does.
• I get trapped in the create → update → resolve loop instead of doing the work itself.
• Some tasks fall into a dead zone: not clearly actionable, often blocked by external dependencies, but also not safe to remove.
• The overdue count keeps increasing, which turns the system into a source of self-blame.
• Tasks are often scoped badly because I underestimate them, which again leads to overdue tasks and self-blame.

The real need is not a better task app. It is someone constantly grooming the task list with enough context and personalisation to keep it aligned with reality. It is closer to having an EA who understands your life, priorities, and goals: someone who can set up, prioritise, reshuffle, and complete tasks on your behalf. The agent should know when to be stern with high-priority or time-sensitive work, and when to be lenient by deprioritising tasks based on context. It can also improve consistency by narrowing the daily scope so that shipping feels possible instead of overwhelming. That is the accountability partner agent I am trying to build.

Setup

The first decision was which task app to use. I tried a few options and eventually settled on Todoist for three reasons:

• It is cloud-based and has good MCP support, so any agent I use on any machine can access it with minimal setup.
• The free version supports unlimited task management, reminders, calendar integration, and excellent tag-based grouping/filtering.
• The app is easy and straightforward to set up on any machine.

The second decision was which agent to use. I chose a hybrid approach. Hermes acts as the main EA: it creates, prioritises, reprioritises, and completes tasks for me. Separately, I use a scheduled monitor in Codex that runs every hour and sweeps through communication channels like Slack, email, and WhatsApp to surface tasks or actionables, then files them in Todoist with context. Codex is better suited for this ingestion pipeline because it has first-class browser and computer-use support. Hermes has an internal browser, but for logged-in websites, Web WhatsApp, and following links to gather context, Codex is currently more dependable.

That said, Hermes is the stronger fit as the always-on EA. It is a self-improving agent running on a cloud machine, connected through Telegram, and available whenever I need it. It can schedule work for itself, handle edge cases, learn my preferences, and extract enough context over time to reduce redundant tasks.

What worked

Once I gave Hermes a clear understanding of my goals and priorities, it quickly set up projects and tasks across a longer horizon. More importantly, it broke them down so each task felt doable in a 60–90 minute window. When more urgent work came up, it was also able to reshuffle priorities while keeping the total load on me roughly constant.

Codex also did a strong job of capturing tasks from my communication channels. Initially it created some false positives - messages that were meant only to inform me, not require action from me. But as I gave it more context, its capture and prioritisation improved. At this point, it rarely creates extra tasks, and very few actionables slip through.

What failed / risks