Agents forget. Your project shouldn't.

A new AI coding session may not have the decisions that shaped the last one.

The agent might not know why I chose Postgres over Microsoft SQL Server, which order I planned to build things in, or why a milestone is waiting on an API decision. If that context isn’t available, I explain it again—or risk getting work that contradicts a decision I already made.

In a company, that knowledge is spread across colleagues, meetings and wikis. As a solo developer, too much of it lived in my head.

One developer, many sessions

Before Workbench, project knowledge lived in a dozen places:

• Decisions buried in chat transcripts.
• Plans in Markdown files that a new agent might not find.
• Progress tracked in my head, drifting away from the code.
• Separate conversations with different tools, each with only part of the story.

More agents meant more output, but also more decisions to record and explain. This brought me back to Req2Code’s original problem: how to keep the intent around the code clear as the work moves forward.

Project Memory: a shared notebook

Each Workbench project has Project Memory: searchable notes for decisions, constraints, preferences, ideas and discoveries. I can write them, and agents with access to the project’s tools can read and update them.

Ideas stay labelled as ideas. Decisions include their source. That distinction matters: a suggestion should not quietly become an agreed direction.

Memory gives a new session somewhere to start. It still needs maintaining, and an agent needs to consult it. The benefit is that the project’s knowledge has a durable home beyond one transcript.

Plan before building

I usually begin in Plan mode, where I discuss an approach with the agent as it explores the project, before implementation. When I’m ready, I move to Build with the appropriate permissions.

Design documents help make an approach reviewable. Saved revisions let me see how it evolved. Project Memory then carries the decisions that other sessions need.

Choosing Postgres over Microsoft SQL Server is one example. After weighing the trade-offs, I recorded the choice and its reasons so later agents could use that decision instead of reopening it without context.

A roadmap next to the code

I also need a plan I can see. I ask an agent to draft a milestone plan from the design documents, then review it as an interactive roadmap in Workbench.

The roadmap brings together milestone dependencies, acceptance criteria, task checklists, blockers, risks and open questions. I can ask the agent to revise it or edit it myself.

Reconcile with code compares the plan with the repository and proposes updates for review. Code inspection can suggest that a requirement is implemented; accepting it still needs evidence appropriate to the criterion.

Progress comes from acceptance and exit criteria marked met with recorded evidence. A date passing or a task checkbox being ticked does not establish that a requirement is accepted.

When I need to share the plan, I can export a slide-style PDF, with an overview and milestone detail. The export reflects the saved plan; its usefulness depends on keeping that plan current.

Picking up where I left off

Recent-work summaries help me return to a project without rereading every transcript. On supported Macs with Apple Intelligence available and enabled, Workbench can generate an on-device recap.

For a broader update, Summarise changes, a Preview feature in this release, helps turn repository changes over a chosen period into a readable report. I review it before using it as a client update, just as I review an agent’s code.

Keep the project’s thread

Sessions end and tools change. Decisions, plans and progress can stay with the project.

That is what I wanted from Workbench: one place on my Mac where I and my agents can plan, build and keep track of software. It helps me direct several agents without being the only place the project’s history exists.

Explore Workbench or download Workbench.

Workbench is made by Req2Code and requires macOS 27 or later on Apple Silicon. AI tools use your own accounts and are installed separately. Apple Intelligence features require a supported Mac and availability in your configuration.