01
Observe
Problems are captured as they happen: the input, the trace, the outcome, and the person who noticed. Recurring failures are clustered so the team sees one problem instead of forty tickets.
Every failure, versioned.
LonsLabs turns what your agents get wrong into verified, reviewed improvements — before any change reaches production.
The fix ships. The lesson doesn't.
A teammate patches the agent at 6pm and the ticket closes. Three weeks later the same failure returns in a slightly different shape, because the fix lived in someone's head and a Slack thread. Troubleshooting solved a problem once. It never became part of the system. LonsLabs exists to close that gap.
Without a learning layer
Fixes evaporate
- Failures live in tickets and chat threads, not in evidence.
- Fixes are applied by hand and nobody re-tests the old cases.
- Prompts, policies, and tools drift apart; nobody knows why a rule exists.
- Two annotators read the same clause two ways and the agent inherits the confusion.
- "Done" means the code shipped, not that the outcome changed.
With LonsLabs
Lessons compound
- Every failure is captured with its trace, input, and human judgment.
- Each one becomes a harness case that every future change must pass.
- The lesson is written into the exact artifact that must change.
- Ambiguities become reviewed, updated guidance — the clause itself is corrected.
- Rollouts carry measurement, and the loop reopens if the number does not move.
The loop
Three stages, one continuous record.
The learning loop sits inside the agent's existing workflow. Nothing moves forward without evidence, and nothing is forgotten once it does.
02
Update the harness
The failure is added to the testing harness. Candidate fixes are run against it and against everything that already worked, so a fix cannot quietly break something else.
03
Deploy
Approved changes roll out with measurement attached. Validation runs in production to confirm the outcome actually changed, not just the code.
Anatomy of a failure
From noticed to measured, in eight steps.
This is the path every failure takes through LonsLabs. No step is skipped, and each one leaves a record the next can verify.
01
Failure noticed
A person or monitor flags a wrong outcome.
02
Captured
Input, trace, and judgment stored as evidence.
03
Clustered
Similar failures grouped into one problem.
04
Harness case
The failure becomes a permanent test.
05
Candidate fix
A change is drafted to the responsible artifact.
06
Verified
Passes the new case and the full suite.
07
Reviewed
Owner and required approvers sign off.
08
Measured
Production confirms the outcome changed.
0
stages in the loop: Observe, Update the harness, Deploy.
0
artifact types a lesson can be written into.
0
record per change, with evidence, owner, and review history.
0
unreviewed changes reach production.
Where lessons land
Pushed into the component that must change.
A lesson only persists if it is written into the artifact that will be reviewed next time. LonsLabs routes each one to the right place and links the edit back to its evidence.
Prompts
The instruction that produced the failure gets the corrected instruction, with the case that broke it attached.
Context
Missing or stale reference material is added where the agent actually reads it, not in a doc nobody opens.
Tools
Tool schemas and guards are updated so the same bad call cannot be issued again.
Policies
Ambiguous clauses get reviewed, rewritten guidance so two people stop reading them two ways.
Evaluation
The failure becomes a permanent test. Regressions surface before rollout, not after.
Environment
Configuration, sandboxes, and data fixtures change alongside the code that depends on them.
SummaryRefund agent escalates when order contains a gift item
Evidencetrace #91f2 · ticket 2231 · 6 replays
Touchesprompts/refund-v14 · policies/gifts.md · eval/refund-suite
OwnerSupport Ops
PermissionsPolicy edit requires Legal · granted
Reviewproposedharness pass 41/41Legal reviewapprovedmeasuring · 7d
Traceability
Every change knows why it exists.
Proposed changes keep their evidence, their owner, and their review history. When someone asks "why does the agent do this now?", the answer is one click away, not one archaeology project.
That record is what turns iteration into accountable iteration: fast enough to run daily, documented enough to satisfy an audit.
- Source-linked: each fix points to the traces and tickets that justified it.
- Permission-aware: edits to sensitive artifacts route to the people allowed to approve them.
- Human-reviewed: nothing ambiguous or unverified ships on its own.
Safeguards
Speed from the harness. Trust from the review.
Four properties every LonsLabs change has, by construction.
Permission-aware
Edits route to the people allowed to approve them. A policy change waits for the policy owner; a tool change waits for the platform team.
Source-linked
Every proposal points to the traces, tickets, and replays that justified it. Reviewers see the evidence, not a summary of it.
Human-reviewed
Anything ambiguous, conflicting, or thinly evidenced is held for a person. Nothing auto-merges into production.
Measured after rollout
A change is not done when it ships. Validation runs in production and reopens the loop if the outcome did not move.
Example
Resolving an ambiguous policy, for good.
Two annotators read the same clause two ways and the agent inherited the confusion. LonsLabs surfaced the conflict, drafted updated guidance, and routed it for review — the clause itself changed, and three new cases now guard it.
404.2 Overlapping speech
41Mark the segment as overlap when speakers talk at the same time.
41Mark the segment as overlap only when two or more speakers produce intelligible words simultaneously for 300 ms or longer.
42Backchannels ("mm-hm", "yeah") under 300 ms are not overlap; tag them as BACKCHANNEL.
43Evidence: 14 disagreements · trace #77a0 · eval/annotation-suite +3 cases
Who it's for
Teams whose agents have to get better every week.
LonsLabs is agnostic to model provider and orchestration framework. It attaches to the workflow you already run.
Support & operations
Refund, triage, and routing agents where one bad decision repeats hundreds of times a day — and where the fix has to be right the first time.
Annotation & data teams
Guidelines that drift, conflicting interpretations, and agents that inherit the ambiguity. LonsLabs turns disagreements into corrected guidance.
Internal platform teams
Shared prompts, tools, and policies that many agents depend on and nobody fully owns. Ownership and review become explicit.
Regulated workflows
Anywhere a change needs an owner, a reason, and a paper trail before it can go live. The record is produced as a by-product of working.
Where it runs
One loop, many kinds of agents.
The same three stages apply wherever an agent makes decisions that matter. These are the settings LonsLabs is built for.
Support & operations
Refund, triage, and routing agents where one wrong decision repeats hundreds of times a day.
Start hereAnnotation & data teams
Guidelines that drift and interpretations that conflict become corrected, reviewed guidance.
Start herePlatform teams
Shared prompts, tools, and policies that many agents depend on get explicit owners and history.
Start hereRegulated workflows
Finance, insurance, and public sector — every change carries an owner, a reason, and a paper trail.
Start hereEvaluation & QA
Each failure becomes a permanent harness case, so regressions surface before rollout.
Start hereTelecom & networks
High-volume assistants where small policy errors compound quickly across channels.
Start herePolicy & compliance
Ambiguous clauses resolved at the source, with the clause text itself updated after review.
Start hereMeasured rollouts
Production validation confirms the outcome moved — and reopens the loop if it didn't.
Start hereQuestions
What teams ask first.
Does LonsLabs replace our agent framework?
No. It is a layer around the agents you already run. LonsLabs observes failures, maintains the harness, and manages the change record; your framework and model provider stay as they are.
Who approves a change?
The owner of the artifact being changed, plus any approvers your permission rules require. A policy edit can require Legal; a tool schema edit can require the platform team. LonsLabs enforces this rather than working around it.
What happens to a fix that passes the harness but fails in production?
The production measurement reopens the loop. The new evidence is attached to the original change, a fresh harness case is created, and the next candidate has to pass both.
Can it handle disagreements about what a rule means?
Yes. Conflicting interpretations are treated as a defect in the rule, not in the people. LonsLabs drafts clarified guidance, routes it for review, and updates the clause text itself once approved.
Where does customer content go?
Traces, prompts, and policies belong to your organisation and are processed only to deliver the service. See the privacy policy for details.
How do we start?
Bring a handful of real failures. We show what LonsLabs would have captured, tested, and changed, on your own cases. Request access to begin.
Get started
Make the next version better than the last one.
LonsLabs is onboarding teams that run agents in real workflows. Tell us where your agents fail and we'll show you the loop running on your own cases.
Request access