The Pearl Success Agent loop now runs overnight

Pearl users who report an issue can now get a first reply from the Pearl Success Agent, without waiting for a person, and the fix can be built while engineers are offline.

In May, Olas introduced the Pearl Success Agent, the agent loop that joins Pearl support and product development in one thread. At the time, every step had a human on it: review the reply, review the scope, review the code, approve the release.

That post expected supervision to narrow as the Pearl Success Agent saw more cases. Four months in, it has. First, replies are designed to go out without waiting for a person. Then, rather than engineers having to write code, they simply review it.

Between May 1 and September 21, 47 Pearl users were unblocked through the loop. "Unblocked" means the user either received steps to resolve the issue on their own, or a fix was released for the bug they reported.

What changed, step by step

Replies

• Then: the Pearl Success Agent drafted a reply, and a human signed off before it went back.
• Now: the Pearl Success Agent investigates the root cause before replying, and for most issues it finds the root cause on its own. The first reply to a new issue is then designed to go straight to the user, and sometimes that reply is enough to resolve the issue. Follow-up replies are still reviewed by a human.

Scope and code

• Then: engineers reviewed every substep of scoping and coding, not just the finished result.
• Now: a scoping agent with access to the Pearl codebase can draft the plan with minimal human input, and one coding agent can implement it while another runs code reviews until the task is complete. Engineers review both before final sign-off.

Reach

• Then: the scoping and coding agents worked only on the Pearl frontend and backend.
• Now: they can now also work on the agents built by Valory: Polystrat, Omenstrat, Optimus, Basius and Pearl Connect. These agents can now go through the same loop, and one engineer can implement work across the whole Pearl stack.

Staying current

• Then: engineers kept the agent's knowledge current by correcting its replies, and every improvement to the loop itself was started by a person.
• Now: after each feature release, the loop aims to update its own knowledge base, and the coding agent can implement improvements to the loop itself.

Prioritization

• Then: prioritization was minimal, so almost every user request reached the build pipeline. Keeping the backlog aligned with actual priorities took a lot of human input.
• Now: the Pearl Success Agent validates each request against the constitution, which also holds the Pearl product strategy. A dedicated prioritization agent (known internally as the grooming agent, after the practice of backlog grooming) then ranks the bugs and features that come out of user issues: bugs by severity, and features by how often users request them and how well they fit the current product strategy. Important bugs are meant to be flagged as high priority, and an engineer is notified to trigger the agent for scoping and implementation right away. The agents help rank and groom the backlog, but nothing moves into implementation without an engineer's approval.

Work that runs overnight

Until recently, the coding agent wasn't given tasks overnight. It had no review loop of its own, so its work needed frequent human checks, and nothing it produced moved forward without an engineer's approval. Leaving it running unattended wasn't practical.

The internal review loop changed that. One coding agent can now implement a task while another reviews it, and the two iterate until the task is complete. Scoped tasks can be handed over at the end of the working day, the agents are set up to work through them overnight, and engineers review the finished work the next morning.

Support had the same limit. Every reply needed a person's approval, so an issue reported at night waited until the morning for a first answer. The Pearl Success Agent now sends the first reply to a new issue on its own.

For users, this means a first reply is designed to arrive without waiting for the working day, and once an engineer starts work on a confirmed issue, a fix can be ready for review the next day.

What stays the same

The constitution still sets the bounds. It is updated as the Pearl product strategy evolves, and requests, scopes and code changes are meant to be weighed against it.

Engineers still review the scope and the code, and a human approves each release before it goes through the normal Pearl release process. When a release lands, the Pearl Success Agent is designed to reach back to the users who reported the issue or asked for the feature.

Step into the loop

Run an agent at pearl.you. To report an issue, contact support from inside Pearl. Each ticket starts the loop.