Agent vault
Every company agent, run under one set of rules.
The vault is where company agents live and run. It decides who can use and change them, keeps them inside your compliance rules, records everything they do, and catches breakages before your people do.
What the vault does
Four jobs, done for every agent.
01 · Governance
Who can use it, who can change it.
Owners, people who can change it, people who can use it. Nothing goes live or changes without the right approval.
02 · Compliance
Inside your rules, by design.
Guardrails, approved models, data region, and evidence ready for your auditors and insurer.
03 · Audit log
Everything on the record.
Every run, tool call, policy block, change and access grant — with who and which data.
04 · Maintenance
Watched while it runs.
Health, versions and checks for every agent. When something upstream breaks, the right people hear about it.
01 · Governance
Clear owners, and nothing changes without them.
Every agent has named owners. Access is set once on the department's parent agent and follows people as they join or leave the team.
- An agent goes live only after an owner approves a test run on real data
- New versions are test-run first; the previous version keeps running until a new one is approved
- Failing checks block approval unless an owner deliberately overrides
- Any agent can be paused with one switch
02 · Compliance
Your rules, applied to every run.
Policies are set once and enforced on every agent, so compliance isn't left to whoever built the workflow.
Guardrails
Block, tag or log.
Stop an agent emailing outside the company, tag personal data in outputs, log every tool call. In plain English, not code.
Models and region
Approved models, in your region.
Choose which models agents may use and where they run — for example EU · Dublin. Customer data is read where it lives and never copied.
Evidence
Ready for the review.
Export the evidence your SOC 2, ISO 42001 and cyber insurance reviews ask for: every run, change and access grant.
Where agents run
Your region. A sealed sandbox for every run.
Choose where your workspace lives, and every agent in it runs there. Each run then happens in its own isolated sandbox, so no agent ever holds the keys to your systems.
- One data region for the whole workspace, such as EU · Dublin. Every agent inherits it, so no one can quietly run one somewhere else.
- Runs, model calls and run history stay in that region. Model calls use approved providers with in-region processing.
- Your data is read where it lives and never copied into Sonsara. Run history is kept for 30 days.
- Each run gets a fresh, isolated sandbox that holds only a short-lived token for that one job. No stored passwords, no connector keys.
- The sandbox reaches your systems only through the agent's approved connectors, and a finished result is never lost, even mid-update.
Fresh and isolated for this one run.
job token only · expires with the runHeld on the agent's account, scoped to what this job needs.
No training on your data.
Run history kept 30 days · your data stays where it lives
03 · Audit log
One record of everything your agents do.
Every run, tool call, policy block, change and access grant is logged with who did it and which data it touched. Run history is kept for 30 days and can be exported at any time.
- Who asked, from which surface — Slack, Teams, Claude or ChatGPT
- Which systems were read or written, and on whose account
- What a policy stopped, and who was told
- Who approved each version, and when
- Done · 41s
Sara M. asked in Slack
Sales agent drafted 3 follow-ups
Gong · readHubSpot · readGmail · draft
- Blocked
Policy · email outside the company
Finance agent tried to email a supplier directly
Gmail · send → vendor.ioblocked
- Done · 1m 12s
Schedule · weekdays 09:00
Supplier reconciliation checked 412 invoices
Xero · readOutlook · send → ap@PII · tagged
- Change
Alex M. approved Supplier reconciliation v2
After a passing test run
Example data
04 · Maintenance
Agents are watched while they run.
Tools change, tokens expire, people leave. The vault keeps an eye on every agent, so a broken one is flagged and its owners asked to fix it — before anyone notices the work didn't happen.
- Health for every agent: when it runs, how often it succeeds, time saved and cost
- Checks on every version, so a change can't quietly make it worse
- A disconnected tool marks the agent down and asks its owners to reconnect in Slack
Gong has disconnected, so Renewal Radar can't run. You're an owner, so it needs you — reconnecting takes a minute.
FAQ
About the vault.
Who can put an agent into the vault?+
An agent's owners approve it after watching a test run on real data. Owners and admins decide who can approve, and you can let department managers approve their own team's agents.
What happens when someone changes an agent?+
The change becomes a new version. It's test-run and checked, and the old version keeps running until an owner or someone who can change it approves the new one.
Is Sonsara SOC 2 certified?+
The vault produces the evidence your own SOC 2, ISO 42001 and cyber insurance reviews ask for about your AI use. Ask us for our current security documentation when you book a demo.
What if a connected tool stops working?+
The agent is marked down, its runs pause, and its owners are asked in Slack to reconnect. The audit log shows exactly what was missed.
Keep reading
Related.
Book a demo
See the AI work already happening in your company.
Read-only access, nothing to install, first findings in days.