TandemLoop is built from the ground up so that external AI agents can work on the same data as equal partners. The app itself contains no AI—you connect an agentic AI tool of your choice (see Connecting an AI Tool). The data folder contains no black-box format, but plain-text Markdown that any tool can read. Still, a clear working agreement applies: you decide what happens to your structure.
The Tandem Principle
The tandem metaphor describes the collaboration well: you sit in front and set the direction; the agent sits behind and provides the power. Neither side acts on its own authority. For the app, this means:
- You define vision, priorities, relationships, and decisions.
- The AI tool keeps track of things, prepares, reminds, recognizes patterns, and suggests actions.
The AI tool is your partner in the tandem—a separate
program, not part of TandemLoop. It creates new items in your system,
maintains tasks assigned to it (assigned_to: ai), and—if
you turn proposals on—suggests changes as proposals.
How Agents Work with the Data Folder
AI agents read the data folder like any other tool—they navigate the folder structure and read YAML frontmatter and the Markdown body. What they may write depends on whether you’ve turned on proposals:
| Proposals off (default) | Proposals on | |
|---|---|---|
| Creating new items: tasks, projects, milestones, areas, folders, views, tags | directly, where they belong; if no place is clear, a task goes into the Inbox | new tasks in the Inbox directly, everything else as a proposal |
Maintaining assigned tasks
(assigned_to: ai): title, body, status (including the
handback to you), date fields and reminder, Waiting For, tags,
delegation, repeat, values of custom attributes |
directly | directly |
| Changing and moving existing items: sorting the Inbox, filling in fields, restructuring projects, adjusting views | directly; on each of your tasks it changes, it leaves an entry in the History | as a proposal, which you review and approve in TandemLoop |
| Deleting, completing or dropping your tasks, changing specs | only at your explicit request; without a request, the agent describes the step in the conversation or asks you a question and waits for your approval | as a proposal |
You rename or delete existing tags yourself in the app. Regardless of
the setting, an agent may create custom attributes and themes and, at
your request, set a reminder on any open task. In both cases, directly
editing _archive/, _system/,
_conflicts/, the app settings in _config/, and
the managed block in AGENTS.md is prohibited.
On every run, the agent records the time in
_config/agent.yaml (last_run). From this,
TandemLoop shows under Back from Agent › With
the Agent when it last worked—see Working Together Autonomously
& Continuously.
Turning Proposals On or Off
Proposals are optional and turned off by default. You can turn them on anytime in Settings:
Settings › AI Agents › Turn on proposals from AI agents
The setting applies to the entire data folder. When it’s off, TandemLoop hides all proposal sections; existing proposal files remain unchanged and reappear as soon as you turn proposals on again.
Proposals—Changes You Approve
With proposals turned on: a proposal is a Markdown file in the
_proposals/ folder. It describes one or more actions (for
example, “Move task X to project Y”), including a rationale. You can
approve, reject, or send back a proposal for revision as a whole or
action by action.
Sidebar › Proposals (at the bottom) › open a proposal › Approve or Reject
You’ll find more about the proposal workflow in Part 4 of the manual.
Custom Attributes—Also Created by Agents
Custom attributes are part of the data model that agents know:
AGENTS.md and _system/docs/data-model.md list
all currently defined attributes with key, type, scope, options, and
description. The app updates this list automatically as soon as you
create, change, or delete an attribute—or an agent writes a new
definition file.
For example, at your request an agent can create a “Client” attribute
for tasks and then fill it in—directly without proposals; with
proposals, directly on its assigned and newly created items and through
a proposal on all others (parameter attributes).
Good descriptions help: Agents read an attribute’s description too. The more clearly it explains what the field is for and how it’s filled in, the more accurately agents enter values.
AGENTS.md—The Instructions for Agents
The file AGENTS.md is in the root folder of your data
folder. It explains the rules to every AI agent: allowed actions,
writing rules, the directory structure, and—with proposals turned on—the
proposal format. The app keeps this file up to date and rewrites it to
match when you turn proposals on or off. You can read it, but you
shouldn’t change the system-managed block between the markers
TANDEMLOOP:START and TANDEMLOOP:END by
hand.
Using agents from outside: When you point Claude
Code, another AI tool, or a script at your data folder, refer it to
AGENTS.md first. Well-instructed agents follow the rules
described there—without proposals they create and change items directly
but delete only at your request; with proposals they submit proposals.
That way you stay in full control.