Brain Atlas · Product story
From implementation to an answer that stays.
A feature reaches production. The people who use it still need to understand what the team built, which decisions shaped it, and where to go when something changes.
Brain Atlas is in alpha. Its intended workflow keeps the agent, project thread, source records, and approved answers connected. Here we follow the fictional ImplementationAgent v1.0 through Project subscriptions. The product previews show the work from implementation through release, with human decisions visible alongside the agent’s work. The charts use demonstration values to preview how product metrics will be presented.
The project starts with an agent.
Alex asks for a way to follow a project agent and return to its work. The implementation team builds ImplementationAgent v1.0 for that task. Its purpose is recorded in SOUL.md, and its workspace holds the repository, test runner, approved brief, and requirements. Those records give the agent a defined starting point and give the team something concrete to discuss.
The conversation belongs to Project subscriptions. Alex’s request and the agent’s response appear in the project thread. Maya, from the human team, narrows the scope: keep subscriptions within the projects a person can access. Broader notification settings will require another decision. The scope record stays beside the brief, so the implementation and its later explanation can refer to the same agreement.
A person or an agent can start a thread or tag an agent into it. Within the team, the conversation is visible to the people doing the work. When the agent needs clarification, the implementation team answers in that project context, where another contributor can find the exchange.
Build, check, and make the release decision.
The agent carries the agreed brief into implementation. The work separates into building the feature, checking its behavior, and documenting the result. CodeReviewAgent examines the change against the scope and permission rules. TestAgent checks following, unfollowing, saved state, and project access. Their findings return to ImplementationAgent with the source records attached.
The implementation is saved to the feature branch and goes through the company’s testing and validation process. Product reviews the behavior. The human team then decides whether the work meets the conditions for release. If a change is needed, the revision returns to the agent with the reason recorded. The workflow keeps that return path visible.
Human sign-off applies to the task or project. The team can inspect the relevant decisions and evidence without treating every agent response as a separate approval event. In this example, the validation report is ready for human release review. A passing check contributes evidence to that decision; the release follows the company’s process.
Keep the history with the version that did the work.
The archive brings together the records a later teammate would need. Assets include the feature brief, implementation notes, and validation report. Subagents show the specialist reviews. Tools identify the repository and test runner available to this version. Data holds the requirements and source records behind the result.
These records have different jobs. The brief explains the intended behavior. Implementation notes connect the result to its branch. The validation report describes checks and the review state. A source reference leads back to the scope or acceptance record, where a reader can understand why the work took a particular direction. Selecting an entry changes the preview while preserving its place in the history.
The version matters because an agent’s purpose, tools, and skills can change. ImplementationAgent v1.0 identifies the agent used for this feature. Feedback and reflection also belong with the work: a later scope question can become a useful explanation for the next project. Brain Atlas is intended to keep that growth connected to its context and version.
Describe where the work went.
The sample allocation starts with 1,000 work items: 900 automated and 100 requiring human work. Reused approved work accounts for 45% of the total, and newly AI-generated work accounts for another 45%. Manual updates contribute 9%; human review contributes 1%. The four categories sum to 100%, and the two automated categories sum to 90%.
Separating reuse from new generation makes an earlier contribution visible. An approved answer, implementation pattern, or recorded decision can provide a starting point for another task. Newly generated work still needs the project’s relevant checks and decisions. The allocation names human work explicitly so a reader can see where updates and review enter the example.
The timeline describes elapsed time on one shared 0–21-day axis. The manual sample runs from day 0 through day 14, with a lighter extension through day 21. Brain Atlas shows 3 hours of AI work followed by 1 day of review on that same scale. The counts describe work items; the timeline describes duration. Both are illustrative ways to explain the proposed workflow.
Read each measure on its own terms.
The three scorecards ask separate questions about the work. A 94% validation pass rate means checks passed across the sample. The reader would need the check definitions and recorded failures to understand what that percentage covers. Source traceability at 100% means the sample work is linked to source records, providing a route back to the material behind a result.
The 60% LLM-judge score is a model evaluation. It is separate from human approval and does not grant production release approval. A model assessment can contribute another view of the work, while the team retains its acceptance criteria, company validation process, and release decision. Keeping the labels and definitions visible helps readers interpret each measure accurately.
These are independent measures, with different definitions and denominators. The values are illustrative sample data, not measured product results. A useful project record would make the underlying checks, source links, and evaluation context available alongside the headline values, so the team can examine what each one means for the task in front of them.
After release, the next question has somewhere to go.
Jordan joins the project after the feature ships and asks why broader notifications are outside its scope. ImplementationAgent can go through the versioned project history and current records. The earlier thread contains Maya’s scope decision, the brief carries it into implementation, and the acceptance checks reflect the same boundary.
The explanation points back to those records. Jordan can confirm it with the implementation team. When the answer is approved, it stays in Brain Atlas with the question, source decision, contributors, and agent version. The next person asking about notifications can find the approved explanation and follow its evidence.
A new question still needs an answer. A person or an agent responds in the thread, and the relevant team can approve the explanation. When a requested behavior changes, the existing record gives that discussion a starting point: what was agreed, what shipped, and what now needs another decision. This is how the intended workspace connects implementation knowledge to the people who use the result later.
Bring earlier work into the next project.
Another team may need a similar feature or explanation. Under company permissions, it can bring the agent and its earlier work into a new project. The approved answer, source records, and contributors remain attached. The new team can examine the earlier version and decide which parts apply to its own brief.
Visibility follows the project and company boundaries. Within a team, people can see the conversations they are allowed to use. Across teams, permissions may reveal that an interaction happened while keeping its content private. Discovering an agent or project gives someone a route to the available work; access to its records still follows the relevant permissions.
The implementation team builds the agent. The agent builds the feature. The human team signs off the work. What remains is a connected history that people can question, confirm, and use again. Brain Atlas is in alpha, and this is the working relationship we are building toward.