TrelloAgents: Two Architecture Patterns Worth Stealing

What got built

In a recent build session, the working group put together TrelloAgents, a small multi-agent pipeline that runs on top of Trello as the shared memory layer. The system worked. More interestingly, the patterns underneath the system are general, and they transfer to use cases that have nothing to do with Trello.

This briefing pulls out the two that are worth stealing.

Pattern one: structured object as agent memory

The claim, AI agents do not need direct communication channels or a shared database. They need a shared, persistent object with the right field structure. Any object with distinct fields for artifact, communication, identity, state, and status can serve as complete agent memory.

In TrelloAgents, a Trello card’s five fields map directly:

• Attachment is the working artifact. One file, replace not accumulate. Prevents version confusion.
• Comments are asynchronous agent-to-agent and human-to-agent communication, fully auditable.
• Description plus hidden HTML metadata carries identity and lineage. The idea_id, iteration, and type persist across every stage.
• List position is the current pipeline state. No status field needed, position is the status.
• Labels are live status flags. In Progress, Approved, Needs Revision, Blocked.

The transferable insight, any structured object you already use can serve as agent memory if the fields are used with discipline. A Notion page. A GitHub issue. A CRM record. A pinned Slack thread with files attached. You do not need to build agent infrastructure from scratch. You need to discipline the use of structure you already have.

The teaching frame, your agents do not talk to each other, they talk through the object.

Pattern two: prompt as configuration

The claim, when agent behavior is encoded in plain text markdown files, it becomes universally editable, separable from infrastructure, and accessible to non-technical owners. This is the most overlooked lever in AI system design.

In TrelloAgents, every agent is defined by a single markdown file in an agents/ folder. Edit the file, the agent’s behavior changes completely. Output format, criteria, persona, domain. No code changes, no deployments, no infrastructure modifications.

The concrete evidence, the entire TrelloAgents pipeline can be adapted for a content marketing use case by editing six markdown files. The Python scripts, the Trello structure, the orchestration, the artifact model, the review-bounce loop, none of it changes.

The transferable insight, separate the behavior layer from the infrastructure layer. The practitioner owns the behavior layer, the prompts. The infrastructure layer is generic and reusable. This is also the difference between a one-time AI project and a maintainable AI system.

The teaching frame, the code is the chassis, the prompts are the car.

Why this matters now

Most cohort members are building AI workflows where behavior and infrastructure are entangled. Separating them is an architectural unlock that makes the system editable by the person who actually owns the work, not just by the person who originally built it.

Both patterns are available today, in any stack, with no new tooling required. The discipline is the unlock, not the technology.