Deputies

Self-Hosted Background Agents for Engineering Teams

The Problem

Coding agents are powerful when a developer is actively supervising them from a terminal. But using agents across an engineering organization requires much more than an LLM and a command-line tool.

Teams need a way to assign work, run agents in isolated environments, preserve progress, share results, manage access, integrate with existing workflows, and review exactly what happened. Building that control plane requires queues, sandboxes, authentication, event streaming, artifact storage, integrations, and reliable recovery.

Without that infrastructure, agent usage remains tied to individual laptops, difficult to audit, and disconnected from the systems where engineering work begins.

What is Deputies?

Deputies is an open-source control plane for delegating engineering work to background agents.

Give it a task through the web UI, Slack, GitHub, or pretty much any system that can send a webhook. Deputies prepares an isolated sandbox with the appropriate repositories and tools, runs the agent asynchronously, and preserves its progress, diagnostics, artifacts, and final response in a durable session.

It gives engineering teams the shared infrastructure needed to move coding agents out of individual terminals and into repeatable, observable workflows.

How It Works

  1. Assign work. Start a session from the web UI or trigger one from Slack, GitHub, a webhook, or a schedule.
  2. Run in isolation. Deputies prepares a sandbox with the selected repositories, environment, tools, and context.
  3. Follow progress. Watch messages, tool calls, diagnostics, generated artifacts, and application previews as the agent works.
  4. Review and iterate. Inspect the result, queue follow-up instructions, download artifacts, or have the agent prepare a pull request.

What's Included

Component Description
Durable Agent Sessions Persistent sessions with ordered messages, searchable history, replayable events, artifacts, and follow-ups
Agent Web UI Assign work, follow live progress, inspect diagnostics, preview applications, and review results
Isolated Sandboxes Run agents with Daytona, Docker, Tensorlake, Kubernetes Agent Sandbox, or AWS Lambda MicroVM environments
Repository Environments Reusable, revisioned work contexts that can include multiple repositories
GitHub Integration Trigger work from GitHub, provide scoped repository access, and let agents prepare branches and pull requests
Slack Integration Delegate work from Slack threads and deliver completion responses back to the originating conversation
Webhooks and Automations Create agent work from external events or scheduled workflows
Artifacts and Previews Preserve generated files, reports, logs, and links to applications running inside agent sandboxes
Queued Follow-Ups Add instructions while work is running or return later and resume the same durable session
Skills and Prompt Snippets Package repeatable team procedures and quickly reuse common instructions
Team Access Control Tenant-wide viewer, member, and admin roles, with support for private sessions
Multiple Model Providers Use standard LLM API keys or supported subscriptions through the Pi agent runtime
Portable Deployment Deploy with Railway, Docker Compose, Kubernetes, or the AWS reference architecture
Observability OpenTelemetry support for control-plane traces and operational metrics

Deployment Options

Deployment Sandbox Provider Best For
Railway Template Daytona or another hosted provider Getting a hosted instance running with minimal infrastructure work
Docker Compose Docker Small teams, a dedicated server, or production-style local deployments
Kubernetes Helm Chart Kubernetes Agent Sandbox Organizations that want agent execution inside an existing Kubernetes environment
AWS Reference Architecture Lambda MicroVM sandboxes Teams standardizing on ECS, RDS, S3, and AWS-native isolation
Custom Containers Any supported provider Organizations with specific networking, security, or infrastructure requirements

Key Design Decisions

  1. Deployability is a core feature. Deputies uses portable infrastructure primitives—Node.js, Postgres, HTTP, containers, and optional S3-compatible storage—rather than depending on a single cloud platform.

  2. Agent work is durable. Sessions, messages, events, artifacts, sandbox state, and callback deliveries survive beyond an individual browser tab or developer laptop.

  3. Sandboxes are provider-neutral. The control plane owns the workflow while execution environments plug in through a stable provider interface. Teams can choose managed sandboxes, Docker, Kubernetes, or AWS infrastructure based on their requirements.

  4. Integrations create work; they do not own it. Slack, GitHub, webhooks, and schedules all feed the same durable session model. Work can be reviewed and continued in Deputies regardless of where it started.

  5. You control the trust boundary. Deputies can run alongside private repositories, internal tools, CI systems, and observability platforms. Operators decide where the control plane, sandboxes, credentials, and model access live.

  6. Start simple and scale when needed. Run the API and worker together for a small deployment, then split and scale web, API, worker, storage, and sandbox orchestration independently as usage grows.

  7. The agent runtime is replaceable infrastructure. Deputies currently uses Pi for real agent work while keeping the durable control-plane concerns (sessions, queues, integrations, artifacts, and sandbox lifecycle) separate from the runner.

Example Workflows

Who Is This For?

Engineering and platform teams who:

Note

Deputies is still early-stage and evolving quickly. Expect new capabilities and occasional breaking changes as the project develops.