Back to Projects
Oct 2026

Agentflow

A portable, skills-first engineering decision router that selects focused guidance based on unresolved coding decisions instead of enforcing a fixed agent workflow.

Overview

Agentflow is a portable library of engineering skills for AI coding assistants. Rather than treating every request as a mandatory multi-agent pipeline, a decision router identifies unresolved questions and selects the smallest relevant specialist set.

Router and specialists

The router begins with repository evidence and the user’s actual goal. It decides whether product behavior, architecture boundaries, implementation, test engineering, code review, security, operations, or another specialist needs to be involved. Deterministic changes can be implemented directly.

Specialists contain explicit responsibilities and bounded review contracts. Their instructions can be installed through the independent skills CLI on supported coding hosts.

Host and distribution boundaries

Agentflow does not supply its own background executor, hosted model, filesystem access, or repository permissions. Authorized host tools perform edits and CI checks. The project also packages skills for ChatGPT Web usage with optional GitHub integration; a generated plugin ZIP alone does not prove directory publication or installation.

Result

The public v0.6.0 release provides a ChatGPT plugin ZIP and the repository exposes the canonical skills library. Engineering outcomes depend on what the connected host can inspect, change, and verify; Agentflow itself is a decision layer rather than a code factory runtime.