Rylvo

Build Agent Evolution

Every failure,
a lesson it keeps.

Agent Evolution reads your agent's replies, finds the failures that keep coming back, and works out why from the traces. It fixes what it can in Test, proves the fix on the replies that failed, and leaves the release to you.

Start your 14-day Pro trialSee how it works

When a failure repeats

The real cause, then a fix that’s checked.

It reads the traces of the failing replies for what actually went wrong, like a limit hit or a tool that was too slow, and offers the fix that matches. You decide how much it does alone.

Ask me

Default

It finds the cause and prepares the fix. You apply it in Test, it replays the failing replies to check it, and you release it.

Fix in Test by itself

When the cause is a setting it can read for certain (tool calls, steps, time, a connector's time limit or retries), it changes it in Test and keeps it only if the replayed replies come back fixed. One fix per bot at a time, five a day at most. A fix that fails its check is undone.

Either way you’re told what happened, and Production changes only in a release you activate. It never writes a rule, loosens a guardrail or touches credentials on its own.

Reviewers

Four specialists read every batch.

Each reads the same replies for a different kind of problem. Every finding comes with its evidence, a likely cause, a suggestion and how sure it is.

  1. 01Pattern

    The same error across replies, the same step failing, loops, and failures getting more frequent.

    pattern
  2. 02Speed

    Steps that are slow again and again, bottlenecks, and the same tool called twice with the same input.

    performance
  3. 03Answer quality

    Empty replies, claims about data that isn't there, answers that contradict the conversation.

    output_quality
  4. 04Input

    Missing checks, tool calls with bad parameters, steps in the wrong order, unhandled errors.

    validation

Rules

Lessons, with a track record.

A finding becomes a rule, written by you or drafted from a review. A rule that’s on applies to every reply and keeps a track record. Once it has been used five times, a rule whose replies keep failing is rolled back.

You set the pace

Watch first. Automate later.

Agent Evolution is off until you turn it on for a bot. Then it only watches. Each step after that is its own switch, so you add more as your team comes to trust what it finds.

  • Record replies

    Keep a record of each reply: the track record every rule and rollback runs on.

  • Spot repeated failures

    After each reply, check whether a failure it has seen before came back.

  • Review on its own

    When 3 of the last 10 replies fail, the reviewers start without being asked.

  • Draft rules on its own

    After a review, write a rule for each finding. It waits for your approval.

  • Turn confident rules on

    A drafted rule above the confidence you set goes live at once. The rest wait.

Questions

Asked often.

Does Agent Evolution change my bot without asking?
Not by default. Agent Evolution is off until you turn it on for a bot. Then it records replies and spots repeated failures on its own; reviewing, drafting rules and turning confident rules on are separate switches, and all start off. When a failure repeats, the default is to prepare the fix and let you apply it.
What does “Fix in Test by itself” do?
When the cause is a setting it can read for certain — the tool-call limit, the step limit, the time limit, or a connector's time limit and retries — it changes that setting in Test, replays the failing replies to check it, and keeps it only if they come back fixed. One fix per bot at a time, at most five a day; a fix that fails its check is undone. You're told either way, and Production changes only when you release.
How do the four reviewers work?
Pattern, Speed, Answer-quality and Input reviewers each read the same batch of replies for a different kind of problem, and return findings with the evidence, a likely cause, a suggestion and a confidence score. Findings point back to the replies and steps that support them. You can edit what each reviewer looks for.
What happens when a learned rule makes things worse?
Each rule keeps a track record of the replies it was applied to and how they went. Rollback is on by default: once a rule has been used five times, it is rolled back if more than 20% of those replies failed. You can change that threshold, and the rule's history stays.
Which traffic does it learn from?
You choose per bot whether Test traffic, Production traffic or both feed logging, detection, reviews and metrics.
Which plans include Agent Evolution?
Agent Evolution is part of Pro and higher, and of the 14-day Pro trial. Reviewer and fix-check model calls run on your own provider key.

Ship it once.
Let it get better.

Start your 14-day Pro trialRead the docs