Manage coding agents.Master complexity.Optimize for cognition.
Agent output scales. Human attention does not.
Use Agent Studio to assign tasks to Codex, Claude Code and Gemini. Keep each task’s scope, project context, runs, evidence and review decisions together.
Connect an installed CLI and a Git repository. Keep execution on configured hosts you control.
- Early alpha
- Self-hosted
- Check setup requirements
Take one small change from request to review
Define the task’s scope before starting an agent. Inspect the changes and evidence, then accept the work or request a focused follow-up.
1. Describe the change
Name the outcome, files or feature in scope, and what should stay the same.
2. Choose an agent
Select a configured CLI and execution host. Start the run and follow its status.
3. Inspect the result
Read the diff and recorded checks. Follow open findings to the affected files.
4. Make a decision
Accept the work or request a specific correction. A failed or incomplete run needs attention.
Inspect an escalated task
This task needs attention. Its review findings, run history and changed files help you decide whether to continue the work, accept the result or stop the task.
Exception example: DEMO-9 is escalated after exhausting its reissue budget. Three findings remain open; develop and main integration are pending.
Keep the request beside the work it produced
Return to the task when you need to understand why a change exists, what the agent tried and which decision comes next.
Scope and context
Attach project rules and information needed for the change. Update them when requirements change.
Runs and evidence
Inspect the run history, diff and recorded checks. Distinguish the agent’s summary from evidence you have reviewed.
Review and follow-up
Record the decision and turn remaining problems into focused work linked to the original task.
Use the agent tools available on your host
Agent Studio coordinates coding-agent CLIs. Models, authentication and usage terms depend on the installed CLI and provider account.
Claude Code
Codex
GeminiShared task state. Local or remote execution.
The Task Server stores tasks, runs and decisions. Agent Runners execute CLIs on configured hosts. Use the browser to review the work.
System map
Central state. Execution anywhere.
Control stays central. Execution lives wherever you allow it.Agent Studio
See and decide in one browser surface.
HTTP client Task Server
Remember and coordinate every task, run and review decision.
authoritative state Agent Runner
Execute coding-agent CLIs and report from a host you control.
agent-runnerMore agent output creates more work to review
When several agents work in parallel, a shared task history helps reconstruct what changed and what needs attention. Start with the next decision and read the evidence it needs. Human review remains necessary.
Six projects with public source code
Alongside Agent Studio, development now includes Runner, Chat, Token Economy, Quality Studio and VOICE. The table shows the published source of each project; token usage comes from local project tracking.
- 6
- public projects
- 5,406
- commits · default branches
- 1.07
- million source lines · incl. tests
- 32.7
- billion tracked tokens · incl. cache
Repository counts use the linked default-branch revisions. Source lines are non-blank lines, including comments; tests are counted separately. Generated files, dependencies, documentation and media are excluded.
Token tracking is incomplete. The figure comes from currently retained usage records across 10 projects and includes cache usage. Runs without saved usage data and separate ad-hoc sessions are not included.
Why I am building Agent Studio
The manifesto explains the motivation: keeping software work understandable as agents take on more of its execution.
Before your first run
Check setup, costs and where work runs, then choose a small change to try.
What do I need to get started?
The documented entry path uses Git and Docker Compose with at least 8 GB free disk space. Add a repository, a configured execution host and an authenticated coding-agent CLI.
Check setup stepsDoes Agent Studio include model access?
No. Agent Studio is open source under Apache 2.0. Configure the CLI, provider account and model access separately. Provider usage and the infrastructure you run can have their own costs.
Understand usage and costWhere do code and agent runs live?
Runners execute CLIs on configured local or remote hosts you control. The Task Server stores task state and run records. What a CLI sends to its model provider depends on its configuration; self-hosting the Studio does not make model calls offline.
Read execution boundariesDoes it replace my editor or coding agent?
The Studio adds task context, run history, evidence and review decisions around your CLI. Continue using your editor and inspect Git changes with your usual tools.
See how the tools fit togetherWhat if a run fails or needs more work?
Inspect the output and changed files, identify the missing outcome or evidence and request a focused follow-up. A finished run still needs a review decision. Agent Studio is early alpha.
Check current limitsChoose one repository and one small task
Connect a repository you can safely change. Inspect the first task’s changes and recorded checks before accepting its result.