When agents break things

Your agents are writing to your database.

Your agents are writing to your database.

Every other layer of your stack has version control. The database has nothing.
It's time to fix that.

Every other layer of your stack has version control. The database has nothing.
It's time to fix that.

Nothing is watching. Nothing is tracking. And when something goes wrong — and it will — you won't know what changed, when, or how to undo it.

Nothing is watching. Nothing is tracking. And when something goes wrong — and it will — you won't know what changed, when, or how to undo it.

We will reply within one business day.

01

The Known

Agents make mistakes.
Thats not the problem.

Agents make mistakes.
Thats not the problem.

Every CTO building with AI already knows this. Agents make unexpected writes. They misinterpret instructions. They act on stale context. They do things at 2am that no human would have approved at 2pm.


That's not a reason to stop building with agents. It's a reason to make sure your infrastructure can handle it when they do. The question isn't will an agent touch your database in a way you didn't intend — it's whether you'll know about it, be able to understand it, and be able to undo it.


Most teams can't. Not because they're careless.

Because the database was never built for this.

Every CTO building with AI already knows this. Agents make unexpected writes. They misinterpret instructions. They act on stale context. They do things at 2am that no human would have approved at 2pm.


That's not a reason to stop building with agents. It's a reason to make sure your infrastructure can handle it when they do. The question isn't will an agent touch your database in a way you didn't intend — it's whether you'll know about it, be able to understand it, and be able to undo it.


Most teams can't. Not because they're careless.

Because the database was never built for this.

02

The Gap

Every layer has a fallback. Except one.

Every layer has a fallback. Except one.

LayerToolCapability
CodeGitBranch, diff, rollback
InfrastructureTerraformState history, plan review
DeploymentsCI/CDStaged rollouts
Secrets & ConfigVault, SSMVersioned, auditable
DatabaseBackups, PITR, replicationRestore, not Rollback. No row level diff. No per-change audit. All or nothing recovery
Databasewith DoltDoltBranch, diff, rollback, audit trail

03

Incident Dermo

An agent updated your risk scores last night.

An agent updated your risk scores last night.

An agent writes to your production database. Something looks wrong — a price is off, a record is corrupted, a table that should have 10,000 rows has 9,847. You notice it 40 minutes later.

Here's what happens next:

An agent writes to your production database. Something looks wrong — a price is off, a record is corrupted, a table that should have 10,000 rows has 9,847. You notice it 40 minutes later.

Here's what happens next:

incident_response.sh — without dolt
INCIDENT OPEN — 4HRS+
incident_response.sh — with dolt
CAUGHT BEFORE MERGE
incident_response.sh — without dolt
INCIDENT OPEN — 4HRS+
incident_response.sh — with dolt
CAUGHT BEFORE MERGE

Your database has none of this.

Dolt CI — the only database with tests built in

You wouldn't let an agent merge code without CI tests. Why let it write to your database withyout them? Dolt runs SQL queries against branches before they merge.

You wouldn't let an agent merge code without CI tests. Why let it write to your database withyout them?

Dolt runs SQL queries against branches before they merge.

SELECT count(*) FROM (SELECT id FROM products WHERE price , 1.0)

Use Infrastructure that makes mistakes survivable

04

The compliance layer

Your auditor has a question you can't answer yet.

Your auditor has a question you can't answer yet.

In a regulated environment — fintech, healthcare, insurance — the agent incident isn't just an engineering problem. It's a compliance problem. When your legal counsel, your board, or your regulator asks "what did your AI touch, and when, and what changed" — your database needs to be able to answer.


Right now, most databases can't. They have logs. They have backups. They don't have version history you can query with SQL

In a regulated environment — fintech, healthcare, insurance — the agent incident isn't just an engineering problem. It's a compliance problem. When your legal counsel, your board, or your regulator asks "what did your AI touch, and when, and what changed" — your database needs to be able to answer.


Right now, most databases can't. They have logs. They have backups. They don't have version history you can query with SQL

REG-01

REG-01

EU AI Act

Article 12 requirements for automatic recording of events ('logs') over the lifetime of the system.

Article 12 requirements for automatic recording of events ('logs') over the lifetime of the system.

Enforcement: Aug 2026

REG-02

REG-02

SEC / FINRA

Model risk management guidelines requiring strict auditability of automated decision systems and data provenance.

Model risk management guidelines requiring strict auditability of automated decision systems and data provenance.

Active now

REG-03

REG-03

Internal Audit

Board and legal requirements for incident response, blast radius determination, and exact remediation steps.

Board and legal requirements for incident response, blast radius determination, and exact remediation steps.

Every incident

Why agentic coding works

Git didn't slow teams down. It made them faster.

Git didn't slow teams down. It made them faster.

Before Git, teams were careful about code because they were afraid of losing it. After Git, they moved faster because the safety net removed the fear. The same logic applies to your database.

Right now, most teams are careful about what they let agents touch because they're afraid of what might happen. That fear is the constraint. Version control removes it.

Your agents can work on branches. Changes can be reviewed before they merge. Mistakes can be undone in one command.

Before Git, teams were careful about code because they were afraid of losing it. After Git, they moved faster because the safety net removed the fear. The same logic applies to your database.

Right now, most teams are careful about what they let agents touch because they're afraid of what might happen. That fear is the constraint. Version control removes it.

Your agents can work on branches. Changes can be reviewed before they merge. Mistakes can be undone in one command.

The question isn't whether you need this.

The question is whether you add it now or after the first incident.

Dolt moves fast,
stay close.

Enterprise

Book an informational session

30 minutes. Your stack, your existing tooling, and how Dolt provides Data security.

We will reply within one business day.

Devs

Join the Discord

Releases, bugs, and the running commentary on whats shipped every week.

Open source · AWS · GCP · Azure · Apache 2.0 · EU AI ACT ready · SQL Compatible

05

HOW IT WORKS

One Connection String
No migration risk.

One Connection String.
No migration risk.

You don't have to rebuild your stack. Dolt replicates from your existing MySQL or Postgres database. Your primary stays exactly where it is. You get a versioned replica — with full branch, diff, rollback, and audit capabilities — from day one.


Or run Dolt natively for the complete Git-style workflow. Same SQL, same wire protocol. Drop-in replacement.

You don't have to rebuild your stack. Dolt replicates from your existing MySQL or Postgres database. Your primary stays exactly where it is. You get a versioned replica — with full branch, diff, rollback, and audit capabilities — from day one.


Or run Dolt natively for the complete Git-style workflow. Same SQL, same wire protocol. Drop-in replacement.

Path 01 — Replicate

Start today. Zero disruption.

Keep your existing primary. Add Dolt as a replica to get immediate audit and rollback capabilities.


→ Zero downtime setup

→ Real data versioned from day one

→ Immediate audit trail

→ Immediate rollback

Keep your existing primary. Add Dolt as a replica to get immediate audit and rollback capabilities.


→ Zero downtime setup

→ Real data versioned from day one

→ Immediate audit trail

→ Immediate rollback

Managed for you

Path 02 — Native

The full safety net.

Run Dolt as your primary database for the complete Git-style workflow with branches and merges.


→ Full Git-style workflow

→ Agents on branches

→ Review before merge

→ Complete provenance from first write

Run Dolt as your primary database for the complete Git-style workflow with branches and merges.


→ Full Git-style workflow

→ Agents on branches

→ Review before merge

→ Complete provenance from first write

Dolt moves fast,
stay close.

Enterprise

Book an informational session

30 minutes. Your stack, your existing tooling, and how Dolt provides Data security.

We will reply within one business day.

Devs

Join the Discord

Releases, bugs, and the running commentary on whats shipped every week.

Open source · AWS · GCP · Azure · Apache 2.0 · EU AI ACT ready · SQL Compatible