Session

From Babysitting Bots to Open-Source Attention Routing: The Colony2 Agent-Cell Model

Overview

ExperienceIn Person
TrackArtificial Intelligence & Agents
IndustryEnterprise Technology
TechnologiesData Marketplace
Skill LevelAdvanced
DOWNLOAD SESSION SLIDES

Single-agent chat loops require a human watching every step, and swarms multiply that burden until coordination becomes the binding constraint. This talk reviews an open-source approach to massive parallel agent coordination that decomposes infrastructure into pluggable components around one primitive: the cell, a repo-scoped runtime with its own context, capabilities and contract. Because agents invoke explicit commands rather than acting in an opaque conversation, every step is an inspectable state transition. The cell’s durable workflow runtime is built on Neon Postgres, recording execution as first-class state: jobs pause at wait points, retry on failure, resume after interruption and carry full audit history. Human approvals become routable primitives in the graph.

Session Speakers

Jacques Nadeau

/Founder
colony2.org

Full Summary

Why AI coding agents exhaust developers, and a new pattern called cellular development

AI coding agents promise more output with less effort, yet many teams feel more scattered and tired. The core reason is an attention mismatch: agents remove much of the keystrokes, but they amplify the highest cognitive load phases where humans make judgment calls. The talk unpacks why today's agent workflows drain attention and proposes cellular development, a sharded, durable way to structure work that respects human focus. The ideas are relevant for anyone rethinking engineering loops in the agent era, including teams exploring Lakeflow Declarative Pipelines and Lakeflow Jobs.

FAQ


A cell is a GitHub repository paired with local memory, explicit guidance and principles, and a durable job history. Cells are scoped by purpose and typically use a single tech stack. Cells can create, split, or merge to keep context bounded and architectures intact.

SWF is spec-first and treats workflow data as central. Jobs, steps, and artifacts are stored durably behind a small REST API set, with operations like create, lease, record step, suspend, and wait. Storage backends are pluggable, for example Neon with object storage, so teams avoid lock-in.

Avoid agents for tasks requiring determinism. They are probabilistic, so a step like running tests may succeed almost always but still occasionally fail to run. Deterministic rules or traditional automation should gate tasks where certainty is mandatory.

Yes. SWF, C2J, example operations, and recipes are open source and positioned as early building blocks. The projects record every step, commit, and artifact to enable inspection, rewind, and collaboration. The author explicitly invited contributors.