An agent to build user-centric agents

For decades, the loop from user to product was slow to close. Requests came in. Roadmaps came out. Features shipped months later. Feedback came back, eventually. Users knew what was missing, but had to wait.

AI tightened that loop. Agents now read a request, understand it, and answer it, running parts of the work that used to wait for humans. Pearl, the Olas ecosystem's flagship product built by core contributor Valory, has been rebuilt around that change.

A Pearl user reached out, worried about losing access to their old computer. The Pearl Success Agent, the agent driving Pearl's loop, took the question and started investigating. Within minutes, it had answered. Within a few days, the Recovery flow via Secret Recovery Phrase shipped, a fallback for users who lose their original setup. The Pearl Success Agent that handled the question reached back to the user who started it.

That's the loop: support and product development on one thread, with the Pearl Success Agent at both ends.

Two speeds, one conversation

The Pearl Success Agent runs at two speeds.

Minutes for a reply. The Pearl Success Agent is designed to read incoming messages, analyze the anonymized logs with knowledge of Pearl's architecture, investigate, and draft a response. A human signs off before the reply goes back.

Hours or a few days for development. When the conversation surfaces a bug or a feature gap, the same thread can kick off a build. Engineers review the scope. Engineers review the code. The work ships through Pearl's normal release process.

Two speeds, on purpose: some conversations resolve in the reply, some become product work. The distance between them is getting shorter.

Both speeds run within a written, version-controlled constitution, a document that defines Pearl's vision and the rules the Pearl Success Agent follows. Scoping decisions, code changes, and replies are weighed against it before shipping.

When a release lands, the Pearl Success Agent is designed to reach back to the users whose requests pointed at it. The user who described the pain hears, personally, that it's gone.

Most software keeps the conversation in one place and the work in another. A ticket here, a roadmap there. Pearl aims to close that gap: for requests that fit, one conversation runs the whole length of the work.

Same path, different surfaces

Several features have shipped through this loop. Each started as a user describing what was missing. Each came back to that user as a feature. Here are three:

• Sidebar rewards indicator: the first feature shipped from user feedback. A glance at the sidebar now shows which agents have earned OLAS activity rewards.
• Backup wallet management: update a backup wallet from inside Pearl, with a notification when it isn't set up.
• Recovery flow via Secret Recovery Phrase: if a computer dies, is lost, or is replaced, the user's recovery phrase serves as the fallback.

Three different surfaces, same path.

What the constitution does

Without a frame of reference, a loop like this would scatter the product across every request that came in. The constitution prevents that from happening.

Valory, a core contributor to Olas, maintains the constitution. It pins Pearl's intent: user-owned AI for a human-first future, so user signals get read against it. User-visible changes are reviewed by a human before shipping.

The Pearl Success Agent doesn't decide what gets built. It hears what users want and surfaces what fits the constitution. The constitution sets the bounds; the loop fills them in.

How each turn tightens the next

Within those bounds, the loop gets better with use as three things come together.

• Resolution coverage grows. Each conversation gives the Pearl Success Agent another resolved case to work from, another shape it can draft on before a human reviews.
• Signal sharpens. When the same gap shows up across users, it earns a place on the build list. Releases are logged in the knowledge base that the Pearl Success Agent draws on, sharpening its read with each turn.
• Engineering capacity opens up. Releases close out the inbound that prompted them, where one exists. The hours that frees go to the next build, sometimes literally the next item in the pipeline.

The human role narrows

It's early. The first Pearl users have already lived through the full loop, describing what they wanted, then hearing it had shipped. The mechanism is in production.

Today, every step has a human on it: review the draft, review the scope, review the code, approve the release. As the Pearl Success Agent sees more cases, supervision narrows. The work it's doing (drafting replies, scoping fixes, writing code) is the work that used to fill the backlog.

What stays is the constitution. What loosens is how often a human has to look at any given step.

Where Pearl is headed: a loop that runs more on its own, bounded by the constitution, with humans reviewing the moments that need judgment.

Co-creation

Olas is built on co-ownership of AI: agents are designed to belong to the people running them, run on their device, and hold their keys. The Pearl Success Agent turns co-ownership into co-creation.

The same conversation that responds to a question can now help reshape the software. The user who flagged the gap hears back when it's closed.

From co-owners to co-creators. The loop from user to product closes faster now.

Run an agent at pearl.you. Step into the loop.