Skip to content

[Discussion] Configurable development pipelines instead of single-agent workflows #4649

Description

@Ic3b3rg

Before submitting

  • I searched existing issues and did not find a duplicate.
  • I am describing a concrete problem or use case, not just a vague idea.

Area

Not sure

Problem or use case

Today a task is typically handled by a single agent from start to finish. While this works well for many cases, different stages of the software development lifecycle often require different responsibilities (planning, implementation, review, testing, documentation).

Currently there is no way to define a reusable workflow where each stage can have its own behavior, provider, prompt, or rules.

Proposed solution

Instead of treating every task as a single agent session, introduce the concept of configurable development pipelines.

A pipeline would represent the software development lifecycle and consist of a sequence of configurable stages (for example: Planning → Implementation → Self Review → Independent Review → Testing → Documentation).

Each stage could define:

  • The provider or model to use.

  • A dedicated system prompt.

  • Project rules specific to that stage.

  • Stage-specific execution settings.

When a stage completes, its output would be passed to the next stage automatically. This would allow projects to encode their engineering workflow directly into T3 Code while keeping every task consistent and reproducible.

The intention is not simply to chain multiple models together, but to separate responsibilities across the development lifecycle and allow each phase to be configured independently.

Why this matters

This would make development workflows reproducible across tasks and repositories, while allowing teams to encode their engineering process directly into T3 Code.

It would also make it easier to specialize each stage of the workflow without relying on a single agent configuration for every responsibility.

Smallest useful scope

Support a simple two-stage pipeline:

Implementation → Independent Review

Each stage should be able to configure its own provider and system prompt.

Alternatives considered

Project Rules partially address this problem, but they appear to customize the behavior of a single agent rather than defining a multi-stage development workflow.

If there is already a recommended way to achieve this with existing features, I'd be happy to learn about it.

Risks or tradeoffs

This feature would introduce additional complexity to both the UI and task execution model. Care should also be taken to avoid turning pipelines into a generic workflow engine.

Another consideration is the overlap with existing features such as Project Rules. If similar behavior can already be achieved today, it would be useful to understand the intended direction before introducing a new abstraction.

Examples or references

Example pipeline:


Planning
  Provider: Claude
  Prompt: Generate an implementation plan.


Implementation
  Provider: Codex
  Prompt: Implement the approved plan.


Independent Review
  Provider: Gemini
  Prompt: Review the implementation, identify bugs, edge cases, and suggest improvements.


Testing
  Provider: Claude
  Prompt: Generate missing unit and integration tests.

This is only an example. The value is not in using multiple models, but in allowing each stage of the development lifecycle to be configured independently.

Contribution

  • I would be open to helping implement this.

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementRequested improvement or new capability.needs-triageIssue needs maintainer review and initial categorization.

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions