Manage coding agents.Master complexity.Optimize for cognition.
Agent output scales. Human attention does not.
Agent Studio turns Codex, Claude Code and Gemini into bounded, reviewable software work. Keep tasks, project context, agent runs, evidence and decisions in one place — so more agent output does not become more mental overhead.
Use the CLIs and subscriptions you already pay for. Run locally or on hosts you control.
- Early alpha
- Build from source
- No hosted service
Agent output scales. Human attention does not.
Coding agents can produce changes faster than a developer can inspect, remember and coordinate them. The bottleneck moves from generating code to maintaining orientation: What is running? Why does it exist? What changed? What can be trusted? What needs a decision?
Human attention and working memory are limited. Context switching, review overload and repeated mental reconstruction make raw agent throughput a poor measure of progress.
Show what needs attention now. Preserve what happened and why.
A workflow designed around the human bottleneck.
Agent Studio organizes coding-agent work around three practical principles: direct execution, bounded complexity and a review surface designed for human attention.
Manage coding agents
Direct Codex, Claude Code and Gemini from one task surface. Choose agent, model, mode and execution host; keep subscription usage visible; compare agents by the work they return.
Master complexity
Turn prompts into bounded, Git-backed tasks. Keep scope, project rules, checks, diffs, screenshots, review findings and follow-up work attached to the same unit of work.
Optimize for cognition
Prioritize what needs human attention. Condense agent activity into reviewable decisions while keeping full context and history available for deeper inspection.
Optimize for cognition does not mean increasing biological capacity. It means designing around limited attention, limited working memory, context-switching costs, review overload and loss of project state.
Works with the coding-agent CLIs you already use
The agents remain the execution engines. Agent Studio provides the shared task, context, evidence and review layer above them.
Claude Code
Codex
GeminiOne task, from intent to decision.
A successful agent run is not completed software. Every run starts with a bounded task, gathers evidence and ends with an explicit review decision.
Shape
Define the intended outcome, allowed scope, relevant context and required proof.
Run
Execute the selected CLI in an isolated Git worktree, locally or on a remote host you control.
Verify
Attach the diff, tests, screenshots, protocol and review findings to the task.
Decide
Accept the candidate, reissue it, split it or create focused follow-up work.
Agents produce candidates. Humans accept software.
See the work, not just the answer.
Intent, run state, protocol, changed files, evidence and the review verdict stay in one surface. Scan the decision first. Then inspect the proof behind it.
One escalated task: review round and open findings above, task state, orchestrator verdict and the task commit side by side below.
Progressive disclosure starts with the current state and required decision, then shows the summarized result. Evidence is directly accessible; protocol, history and diffs remain available for deep inspection.
Central state. Execution anywhere.
One Task Server keeps every task, run and review decision coherent. Agent Runners execute coding-agent CLIs on any host you control. Agent Studio gives the work one browser surface.
Agent Studio · See and decide
Client
Task Server · Remember and coordinate
Authority
Agent Runner · Execute and report
Execution
Control stays central. Execution lives wherever you allow it.
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 clientTask Server
Remember and coordinate every task, run and review decision.
authoritative stateAgent Runner
Execute coding-agent CLIs and report from a host you control.
agent-runnerA working reference — not a demo deck.
Agent Studio is its own first reference. In the April 18 – July 7, 2026 build snapshot, one developer recorded 1,950 commits, 275k product lines and 163k test lines in about eleven weeks. The point is not volume alone: the work remained divided into bounded tasks with context, evidence and review history.
- 1,950
- commits · Apr 18 – Jul 7, 2026
- 275k
- product lines · historical snapshot
- 163k
- test lines · historical snapshot
- ~11 wks
- one developer · historical snapshot
Repository snapshot as of 2026-08-09. Commit totals are GitHub default-branch commit counts. Product and test figures count non-blank lines in checked-in source files on main; test directories and .spec/.test/Test files count as tests. Generated, vendor, build, coverage, documentation, public-asset and .quality paths are excluded.
The future of software development will not be decided by who produces the most code, but by who masters complexity best.
Agent Studio keeps fast agent work bounded, contextualized, inspectable and under human control.
Start with one reviewed task.
Clone Agent Studio, connect a safe repository and take one deliberately small change from bounded intent to an evidence-backed decision.
Your first success is not a green agent status. It is a task you have actually reviewed.
- Early alpha
- Build from source
- No hosted account