01
Capture bot runs
Build a usable record of what happened.
Agent Evolution logs the run outcome, steps, tool usage, latency, errors, triggered guardrails, outputs, and which learned rules were applied.
Agent Evolution
Turn bot failures into safer improvements.
Detect recurring failures, review the evidence with four specialist agents, and turn approved findings into versioned rules—with configurable automation and rollback.
4specialized reviewers
6rule targets
2traffic environments
1versioned rulebook
How Agent Evolution works
From production evidence to a measured improvement.
Agent Evolution is not an autonomous prompt-rewriting black box. It is a configurable pipeline that detects recurring problems, asks specialist reviewers for evidence, turns approved findings into versioned rules, and measures what happens next.
A controlled improvement loop
PER BOT · CONFIGURABLE · AUDITABLE
Run logs
steps · tools · outcomes
Failure patterns
clustered · severity-ranked
Reviewer insights
evidence · confidence
Versioned rules
approve · measure · rollback
Manual approval by default
Rules reach the next run when active
Regression rollback is configurable
02
Group recurring failures
Separate patterns from one-off noise.
The detector groups repeated error signatures, timeouts, empty or repeated responses, parsing failures, guardrail blocks, and loop behavior into severity-ranked patterns.
03
Review the evidence
Analyze runs from four distinct angles.
Pattern, Performance, Output Quality, and Validation reviewers inspect the selected run batch and return evidence-linked findings with confidence scores.
04
Propose a rule
Turn a finding into an actionable change.
Create a candidate rule from an insight, author one manually, or let the configured pipeline generate candidates. Similar suggestions are deduplicated and scored.
05
Approve and apply
Keep a person in the loop—or automate by policy.
Manual approval is the default. When enabled, qualified candidates can activate automatically and active learned rules are injected into subsequent bot runs.
06
Measure and roll back
Check whether the rule actually helped.
Application, success, and post-rule failure counts feed effectiveness scoring. Regression protection can roll a rule back after enough negative evidence, while version history remains available.
Four specialist reviewers
One run batch, four useful perspectives.
Each reviewer has a specific job. Their findings include the issue, likely root cause, a concrete suggestion, confidence, and the runs and steps that support it.
01
Pattern
Recurring behavior
Finds repeated errors, failing steps, loops, and increasing failure frequency across runs.
02
Performance
Latency and waste
Finds slow steps, bottlenecks, redundant tool calls, oversized context, and abnormal run time.
03
Output Quality
Answer quality
Flags empty, inconsistent, unsupported, contradictory, or question-ignoring responses.
04
Validation
Logic and inputs
Checks missing validation, bad tool parameters, unhandled edge cases, and step-order problems.
Reviewer calls use the model selected in your bot’s Evolution settings and your configured provider key.
Control before autonomy
Decide exactly how far it can go.
Enable the pipeline per bot, start with observation only, and add automation when your team is comfortable. The automatic stages are separate controls—not one irreversible switch.
Choose the evidence
- Production and test traffic scope
- Manual or threshold-triggered reviews
- Configurable run and failure thresholds
Choose what automates
- Automatic logging and failure detection
- Optional automatic reviews and rule generation
- Manual approval or automatic rule activation
Choose the guardrails
- Confidence threshold and active-rule cap
- Regression threshold and automatic rollback
- Per-bot enablement and reviewer selection
Recommended starting point
Observe → review → approve manually
Turn on logging and failure detection first. Run reviews when enough evidence exists, inspect candidate rules, then activate only the fixes you trust.
Auto logging: onFailure detection: onAuto review: optionalAuto activation: off
What your team can do
A workspace for findings, not another mystery score.
The dashboard uses real per-bot data. Inspect run logs, review findings, resolve recurring failures, manage candidates and active rules, compare versions, and tune the pipeline from one place.
AGENT EVOLUTION / SUPPORT BOT
ONReview & Insights 6Failure Patterns 3Rulebook 8
OUTPUT QUALITY
Responses omit required next steps
0.88 confidence · 7 affected runs
Propose fixDismiss
PERFORMANCE
Repeated customer lookup adds latency
0.81 confidence · 5 affected runs
Propose fixDismiss
Require a clear next step before closing
Candidate · output filter · version 1
Overview
Success, failures, patterns, insights, reviews, and active rules.
Run logs
Filter by outcome and environment, then inspect steps and output.
Improve
Move from reviewer insight to failure resolution and rule activation.
Configure
Set thresholds, reviewers, environment scope, memory, and automation.
Agent Evolution FAQ
Clear answers before you enable it.
The most important operational and product questions, without pretending every team wants the same level of automation.
Does Agent Evolution change my bot automatically?
Not by default. Logging and detection can run while rule generation and activation stay manual. Each bot has separate switches for automatic review, rule generation, and rule injection.
What data does a reviewer inspect?
The selected batch can include inputs, outputs, run steps, tool usage, errors, guardrail results, latency, tokens, and the rules applied to those runs. Findings point back to affected runs and steps.
Can I review test traffic separately?
Yes. Evolution supports production and test environment scope, so teams can decide which traffic feeds logging, detection, reviews, and metrics.
Can my team edit or create rules manually?
Yes. The Rulebook supports manual rules, candidate activation, disablement, version inspection, and rollback. Access follows workspace roles.
What happens when a learned rule performs badly?
When regression protection is enabled, the system tracks post-application failures and can roll back an active rule after the configured threshold and minimum evidence are reached.
Which plan includes Agent Evolution?
Agent Evolution is available on Pro and higher plans. Reviewer model calls use your configured provider key.
Included on Pro and higher
Improve from evidence—not guesswork.
Enable Agent Evolution for one bot, keep approval manual, and see what the first review finds. You can add automation later.