AI Agents

AI Agents · By Agent-Pilled

AI agents for work that needs review.

Open-source agents that complete tasks through independent review. Being built for Microsoft Teams: mention an agent, give it a task, and get a reviewed result using your chosen tools, models and providers.

Discuss a pilot Read the design

Free Apache-2.0 core · Independent review · Self-hosted

Planned workflow

A task, a result, a separate review

  1. You

    Mention an agent in Teams

    Describe the task and the result you need. Only authorized people can start work.

  2. Author

    Produces the work

    Uses the context and tools it is permitted to access, in an isolated environment.

    Own credentials · own sandbox

  3. Reviewer

    Checks that exact result

    A separate agent checks the work against the task. The author cannot approve its own result.

    Own credentials · own sandbox

    Findings go back to the author. A revised result needs a new review.

  4. You

    Approve where required

    Your rules decide which actions need your approval. The reviewed result returns to the thread.

Review belongs to a specific result: the exact commit for a code change, or a version of a document or report. Revising the work requires a fresh review.

Tools, models and providers

Choose what runs each role

The first coding workflow targets Claude Code. Codex and Pi adapters are planned. Changing the tools should leave the role’s permissions and required reviews intact.

  • Agent tools

    The program that runs the agent, and the integrations it can use. Claude Code is the first harness target, with Codex and Pi to follow. Teams and access to non-code work are planned separately.

  • Model

    Configure the model per role. An author and a reviewer could use different models, including open-weight models through the planned Pi adapter. Each combination needs evaluation.

  • Provider

    Choose a supported service or endpoint for the model. The planned Pi adapter provides a path to multiple hosted providers and self-hosted models.

Supported combinations will depend on the adapter, provider and sign-in path. No model pairing has been established as best.

Initial scope and broader direction

Start with software delivery. Extend to other work.

  • Code changes

    Take a GitHub issue through implementation, independent review and merge, with review tied to the exact commit.

  • Reports

    Assemble delivery reports from tickets and pull requests. Review the report’s sources and conclusions before returning it to the team.

  • Documents and other tasks

    Produce documents and other results with a separate reviewer, task-specific checks and your approval rules.

Open source and enterprise

Free core. Planned paid enterprise capabilities.

  • Review controls in the free core

    The core stays free under Apache-2.0, including role isolation and the review and approval gates. Model usage and hosting are separate costs.

  • Self-hosted deployment

    Run the software on infrastructure you control. Credentials for model access stay separate from each role’s credentials for the systems it works in.

  • Organization capabilities

    Paid additions are planned for administration, audit and reporting, policy deployment and support.

Development status

What exists, and what comes next

The first GitHub workflow is in development. Teams integration and other workflows follow. The source is in agent-pilled/ai-agents; the initial coding runtime uses the name Forgecrew.

  1. Design Done

    A GitHub delivery design, glossary, decision records and dated research.

  2. Runtime groundwork Done

    CI, fake and GitHub forge adapters, GitHub App manifests and a sandbox experiment.

  3. GitHub delivery Next

    Review gates, the author workflow, retries and capacity controls. Claude Code is the first harness target.

  4. Microsoft Teams Planned

    Start tasks by mention and return reviewed results to the conversation, with appropriate access controls.

  5. More work and tools Planned

    Reports and documents, Codex and Pi adapters, and review controls for each supported kind of result.

  6. Enterprise Planned

    Organization administration, audit, reporting, policy deployment and support.

FAQ

Questions

Can I use it in Microsoft Teams today?

Teams integration is on the roadmap. The first release focuses on GitHub software delivery. Discuss a pilot to help define your team’s requirements.

Is this only for coding?

Coding is the initial workflow. The product direction also covers reports, documents and other tasks that can be checked against a defined result. See the development status for the current scope and planned extensions.

What makes the review independent?

The author and reviewer have separate authority, credentials and isolated environments. The coding design binds a required review to the exact commit. The planned non-code workflows must likewise bind review to the specific result version, so a revision cannot keep an older acceptance.

What does it cost?

The core is free under Apache-2.0. Model usage and hosting are separate costs. Enterprise capabilities are planned as paid additions. The technical design covers intended subscription and API sign-in paths.

Why does the runtime use the name Forgecrew?

Forgecrew is the initial coding runtime name. The product is Agent-Pilled AI Agents.