Learnings

Agents that get better every day.

Learnings turns every run into signals, finds what's going wrong across your fleet, and proposes the fix. Approve it once, and every agent follows it from the next run.

Book a demo

Learnings · Support Studio · Last 30 days

Illustrative

Agentic health

94%

+6 pts this month

4,212

Runs analyzed

5

Learnings ready

12

Rules added

ComplianceReliabilityLatencyCostRule failures
1
Rule failing7 of 20 runs

Refund policy not cited before approval

2
Redundantadds latency and cost

18 rules never fired in 60 days

3
Driftimpact high

Cost per run 41% above this agent's baseline

4
Missing rulerecommended

Escalate refunds above the cap to a person

5
Reliabilityimpact low

Lookup tool timing out on 1 in 12 runs

Selected learning

Refund policy not cited before approval

Root cause

The refund policy section isn't in the agent's context.

Proposed fix

Add section 4.2 to the agent's context and require a citation.

Evidence

Run 398Run 401Run 405+4
Flag for fixAdd ruleCode fixChange contextWatch

Apply to

This run
This agent
Every agent in the Studio
Approve

Why it matters

Your agents run faster than anyone can review them.

Review coverageToday

Your team reviews about 1 in every 100 runs.

ReviewedNever reviewedThe other 99, nobody looks at.

AISquare · Learnings

Refund agentLiveRun history · Last 40 runs
40 runs
Without Learnings

Fixed by hand 3 times. Still comes back. Run #324 knows no more than run #1.

With Learnings
Approved

Pattern found at run 7. Approved once. Gone from every agent after.

Without Learnings

  • 10 mistakes
  • 3 fixes by hand
  • Still recurring

With Learnings

  • 1 mistake
  • 1 approval
  • 0 repeats
MistakeFixed by handApprovedClean run

Observability shows you what went wrong. Without Learnings, someone still has to find it, fix it and remember it, run after run.

How Learnings works

Detect. Decide. Deliver.

01 · Detect

Find the patterns one run can't show.

Learnings reads every run and groups what keeps going wrong, so you see causes, not alerts.

  • Rules that keep failing, with the root cause
  • Rules that never fire, adding latency and cost
  • Cost and speed drift against each agent's baseline

AISquare · Learnings

Support Studio/Learnings

Last 30 daysAll agentsOpen · 5
LearningCategoryImpact

Refund policy not cited before approval

Rule failing

18 rules never fired in 60 days

Redundant

Cost per run 41% above baseline

Drift

Escalate refunds above the cap

Missing rule

Lookup tool timing out

Reliability

AISquare · Review

Proposed learning

Always cite the refund policy section before approving a refund.

Action

Flag for fixAdd ruleCode fixChange contextWatch

Apply to

This runThis agentEvery agent in the Studio

Evidence

Run 398 · 401 · 405 +4

Reviewer

Support lead

Status

Pending approval
Approve and add to rulebook

Kept on the record

02 · Decide

Nothing changes until your expert says yes.

Each learning comes with evidence and a proposed fix. Your expert chooses what to do and how far it goes.

  • Flag, add a rule, change context or send a code fix
  • Apply to one run, one agent or every agent in the Studio
  • Every decision is attributed and kept on the record

03 · Deliver

The rule applies from the next run.

Approved fixes join the rulebook with no retraining or redeploy. Engineering gets a brief it can act on.

  • Enforced on the next run, at the scope you chose
  • A repair brief for your developer or coding agent
  • The same insights over the API and the learnings MCP server

repair-brief.md

repair-brief.md
Markdown

## What's happening

A rule fails on 7 of the last 20 runs.

## Why

The agent doesn't have the policy section in context.

## Suggested fix

Add section 4.2 to the agent's context.

## Evidence

Runs 398, 401, 405 and 4 more.

insights API
curl -s -H "X-API-KEY: $EXPLAINABILITY_API_KEY" \
  "$EXPLAINABILITY_GATEWAY_URL/v1/studios/$STUDIO_ID/insights?window=20"

The same insights are available to your tools. Read the docs

Built for enterprise

Governed by design.

On the record

Every learning, approval and rule change is attributed and kept.

Human approval

Nothing reaches your agents without an expert's sign-off.

Your rulebooks

Approved fixes land in the rulebooks you already run.

Scoped and screened

Learnings stay in your workspace, with sensitive data screened before anything is saved.

Why it matters

Finding issues is the easy part.

Dashboards

Show you what went wrong. You still have to read them.

Issue trackers

Suggest a fix. Someone still has to apply it by hand.

Learnings

Turns an approved fix into a rule the agent follows from the next run.

FAQ

Questions about how agents learn and improve.

Does Learnings change my agents automatically?

No. Learnings proposes fixes. Nothing changes until one of your experts approves, and every approval stays on the record.

How does a learning become a rule?

When a pattern repeats across runs, Learnings proposes a fix with the evidence behind it. Once your expert approves, it joins the rulebook and the agent follows it from the next run, with no retraining.

What does it watch?

Cost, speed, reliability and rule results on every run, compared with each agent's own baseline over 7, 30 and 60 day windows.

How far can a fix go?

You choose: just that run, that agent, or every agent in the Studio.

How is this different from evals?

Evals score outputs. Learnings finds the patterns behind them and turns an approved fix into a rule your agents follow.

What is agentic health?

One score per agent that rolls up rule failures, drift, reliability, redundant rules and missing rules, so you can see at a glance which agents need attention.

Can a coding agent use it?

Yes. Learnings writes a repair brief a developer or coding agent can act on, and the learnings MCP server exposes the same reads to Claude Code, Cursor and any MCP client.

Is it available on every plan?

Learnings is available to all AISquare customers.

Stop making the same fix twelve times.

Start your journey with AISquare

Connect an agent and let every approved fix make the next run better.