Raz Berman Gold

Raz Berman Gold is co-founder and chief technology officer of Overcut, where she is building systems that coordinate AI agents across software delivery. Her work addresses what happens after an agent can write code: how an organization assigns work, checks results, retains useful lessons, and keeps control over what its agents may do.

From Microsoft engineer to founder

Raz spent a decade at Microsoft, joining as a student and growing into an engineering leader responsible for roughly 20 engineers. Her work ranged from shipping Surface Pen features to helping turn Video Indexer into an Azure product and running AI services at the edge. That progression took her from individual product contributions into the operation of AI services and the leadership of the people building them.

She came to value developing engineers alongside developing software. When she left Microsoft, she singled out the satisfaction of watching colleagues grow into leadership roles. Leaving was a deliberate move toward building something of her own, after years at a company and with a team she valued.

At Overcut, her co-founders are chief executive Yuval Hazaz and chief revenue officer Oren Alaluf. Together, they are building infrastructure for engineering work that begins with events in existing tools and continues in the background. A ticket can trigger repository selection, implementation in an isolated sandbox, automated review, a human approval step, and the creation of a pull request. Progress returns to the ticket rather than requiring everyone to follow a separate agent conversation.

Raz’s argument for organizational automation starts with a practical constraint: a powerful coding assistant still consumes a developer’s time if every task requires someone to prompt it and direct the next step. She expects engineers’ responsibilities to shift toward delegating work, supplying context, reviewing results, and designing the processes around agents. She also questions whether every company should build its own orchestration platform. Her preferred division leaves teams concentrating on their particular knowledge and workflows while a shared platform supplies the machinery for running them.

Learning without spreading mistakes

In her writing, Raz examines workflow memory: retaining lessons from completed agent runs without allowing every observation to become an instruction. In her writing on self-improving AI, she uses a familiar failure to explain the need. An agent discovers how to handle a troublesome integration test, finishes its task, then loses that knowledge and repeats the investigation on the next ticket. Saving the lesson can help; saving a mistaken explanation can make future runs worse.

The Overcut approach she describes places much of the engineering burden on deciding which lessons deserve trust. The mechanisms she describes are team-built Overcut features, with explicit tradeoffs:

• Learning scoped to a workflow step. A lesson stays attached to the step where it arose. Guidance useful during code review does not automatically reach planning or other workflows. Raz accepts slower dissemination in exchange for containing the damage from a confident but incorrect lesson.
• Tentative memories before active guidance. A single observation can be recorded without immediately influencing agents. Repetition in a separate run can promote it to active use, while a person can confirm it sooner. This gives unusual outcomes a place in the record without letting them immediately redirect future work.
• Asymmetric trust. Harmful memories lose weight faster than helpful ones gain it; unused memories decay, and weak ones are archived. The design favors withdrawing misleading advice promptly over preserving every potentially useful lesson.
• Separate evaluation and human approval gates. Workers can write memories, but a separate retrospective examines executions and adjusts their standing. People retain control over permissions, required approvals, and completion criteria. Learning can change an agent’s approach without granting it new authority.

Overcut’s retrospective system makes those changes inspectable: each retrospective produces a run showing its findings and memory updates. This matters because a team needs to understand why an agent behaves differently today than it did last week.

Raz also identifies the limits of the approach. A major refactor can invalidate an old lesson without an obvious automatic signal. Connecting a successful outcome to one particular memory remains an estimate. And an evaluator that reads engineering traces is itself an agent, rather than an infallible judge. Her position leaves those problems visible: useful self-improvement requires readable changes, restricted scope, and checks that the workers cannot rewrite.