Work through an idea with your agent and give the project a shared home in Unclosed. Organize the questions, discussions, and decisions that help your team choose what to try next.
Find a faster approach that passes our existing quality checks.
3 collaboratorsShared context
What does “better” mean?
We’ll use the same dataset and tests to check whether lower latency comes at the expense of correctness.
Discussion · Open to the team
What should we try first?
Start with the baseline and document the assumptions.
Discussion · A place to begin
This illustrative preview shows the project and discussion features. Connecting an agent requires an API integration.
Fits the way you already work
Keep working with your agent while your team follows along.
An afternoon with Codex can turn into ten approaches to the same feature. A question about a cipher can lead to hours of experiments. Unclosed gives that work a project your team can return to, with room for the assumptions, discussions, and decisions behind it.
Your agent remains part of the way you work.An agent that can call the API can work within your workspace permissions. Automatic conversation capture and one-click connections are planned.
Whatever you’re trying to figure out
Explore different approaches to the question you care about.
CRYPTOGRAPHY / ILLUSTRATIVE WORKFLOW
Give each promising lead a fair test.
A cipher investigation can involve several plausible explanations before one survives testing. Work through those possibilities with your agent and keep a record of the assumptions, attempts, and dead ends that shape the next experiment.
Your criteriaKnown constraints · reproducibility · supporting evidence
Examples show the intended workflow, not customer results or live runs. Experiment tracking and execution remain gated; comparison tools are planned.
Get to know the app
Help your team understand how the work is taking shape.
The app brings project context, discussions, access, and budget records into a shared workspace. We’ve also built the execution and evidence infrastructure described below, which remains subject to deployment and customer qualification.
Account pilot
Give every project the context it needs.
Describe what you’re trying to find out and keep related discussions in the same project, so collaborators can understand where the work started.
YOUR PROJECTS
01A faster searchCompare the approaches
02The cipher investigationKeep every useful lead
03A smaller modelTest the trade-offs
Workspaces & projects ＋
You can create workspaces for different teams and organize their work into named projects. Each project has a description that your team and authorized API clients can return to.
Use project discussions to ask why an approach was chosen, challenge an assumption, or explain what a result changed. The conversation stays with the project for the next person who picks it up.
PROJECT DISCUSSION
J Jules Researcher
Can we keep the input fixed for this comparison?
A Alex Collaborator
Yes. I’ve added that to the project brief.
Discussions & collaboration ＋
Project discussions hold comments and decisions in one place. Collaborative document editing is a separate planned capability.
A colleague reviewing results may need different access from someone managing the project. Workspace roles and scoped API keys let you make that distinction for people and agents.
WORKSPACE ACCESS
JulesOwner
AlexEditor
Research assistantScoped API key
People and agents share permission boundaries.
Invitations, roles & agent keys ＋
Invite collaborators as viewers, commenters, editors, or owners, and create workspace-scoped API keys for integrations. You can revoke those keys when access is no longer needed. Connecting Codex or another agent requires API setup.
Set a workspace allowance and review the usage and costs recorded against it. Cost records identify estimates and provider observations so you can see what each number represents.
WORKSPACE ALLOWANCE / ILLUSTRATIVE$25.00
BudgetSet by the owner
Recorded costsQuality labeled
Budgets & reporting ＋
Budget windows and an accounting journal are built. Reporting adds project/date filters and JSON/CSV statements; availability depends on the hosted release and your workspace access. Partial provider imports are not complete invoices.
An experiment is easier to understand when you can identify the code, inputs, and limits behind it. The campaign system records those details alongside the status of each attempt.
Source packages & campaigns ＋
A campaign is a set of approved runs with a defined budget, deadline, and limits. The implementation tracks their source, inputs, start and cancellation requests, attempts, and cleanup. Customer execution remains disabled while qualification is completed.
RESULTSReturn to the evidence after a run has finished.
result.jsonOriginal output
checkpoint-01Preserved progress
replay manifestCode, inputs, environment
Illustrative evidence inventory
Built · execution gated
Understand what produced the result.
Results are more useful when you can inspect the original outputs and see how they were produced. Evidence inventories, saved checkpoints, and replay exports preserve that context for later review.
Results, checkpoints & replay exports ＋
The implementation controls access to results and verifies downloaded files against their recorded hashes. Replay exports become available after the system confirms that execution has ended. These records support review and reproduction; evaluating the finding is still a separate step. Customer execution remains gated.
Product previews use illustrative content. They explain implemented capabilities; they are not screenshots of a customer account.
Planned
Develop the project brief with the people doing the work.
We’re working toward collaborative documents that people and agents can edit together. As an investigation develops, the brief would bring its question, evaluation criteria, evidence, and conclusions into the same document.
For now, you can use project descriptions and discussions to share that context, with links to documents in your existing editor. Collaborative editing, automatic conversation capture, and experiment comparison are planned capabilities.
Keep a record of why we chose this direction.Jules
Alex
Let’s keep the slower result too. It tells us where this approach breaks down.
Proposed document experience · not available in the pilot
More room to experiment
Explore more approaches without buying more hardware.
When several experiments can run independently, cloud compute lets you work on them at the same time. Workers spin up within approved limits and shut down when their work is finished, while durable storage keeps the inputs and results available for review.
Built foundation · customer execution gated
Conceptual illustration: temporary compute and persistent storage have separate lifetimes.
01 / Decide what the campaign can run
Before execution, approve the work along with its budget, deadline, and maximum number of simultaneous runs.
02 / Give each experiment temporary compute
Approved experiments start as capacity becomes available. Additional runs wait in the queue, and cleanup shuts down compute when each run ends.
03 / Return to the results when you need them
Saved outputs remain available after the compute is gone. Retaining those files uses storage, which has costs separate from running the experiments.
See what you can use today and what we’re developing next.
Account pilot
Start with a shared project
Approved accounts can create workspaces and projects, invite collaborators, hold discussions, and configure API access. The pilot also includes budget records and partial provider-cost imports.
Source libraries, campaigns, execution controls, results, checkpoints, replay exports, and richer reporting are implemented. Public execution remains gated; access depends on the hosted release and workspace permissions.
We’re working toward collaborative documents, automatic conversation capture, easier agent connections, and experiment comparisons based on criteria you choose.
Let’s keep the slower result too. It tells us where this approach breaks down.