Use case · Cross-industry
Agents that watch, diagnose and restart, under your rules
A run fails at 3 a.m.: an agent reads the logs, finds the cause, suggests a fix and restarts the job. It acts through a service account scoped to one project, sensitive actions wait for your approval, and every call shows up live in the console.
The Agents panel: every connected agent, its permissions, and each of its actions, live.
Real graal console interface, not retouched; fictional demonstration data (tenant energie-demo — people, projects and tokens are invented).
The challenge
Busy on-call rotations, agents to keep in check
- Dozens of runs every night, and failures that wait until morning.
- Repetitive diagnoses: a late file, a column that changes type, a quota reached.
- Promising AI agents that cannot be left to act with administrator rights.
- The need to know, afterwards, who did what.
How graal handles it
An operations agent, under control
Step 01
Connect
Your agent — Claude, an agent built into graal or any MCP client — connects to graal's MCP server with its own service account, scoped to one project.
Step 02
Set the rules
A per-agent policy sets the allowed operations, projects, a maximum cost and time windows. No delete tool is enabled by default.
Step 03
Diagnose
The operations agent reads the logs, events and metrics of a failed run, then writes a diagnosis the team can read.
Step 04
Act
Restart, fix a parameter, reschedule: routine actions go through, sensitive ones wait for approval by someone other than the agent's creator.
Step 05
Trace and revoke
Every call shows up live in the console and stays in the audit trail. Revoking the service account cuts the agent off.
The typical scenario: a data team of a few people, a fleet of nightly jobs, and an on-call rotation that should only be woken up when it matters. You start read-only — the agent diagnoses and suggests — then open up, one action at a time, what it may do on its own.
On 09/12/2025, Anthropic handed the MCP protocol over to the Agentic AI Foundation, a directed fund under the Linux Foundation.
What you get
- Failures diagnosed before the team arrives
- Routine restarts done without intervention, the others submitted for approval
- A log of every agent action, readable and exportable
- Agent permissions you widen at your own pace
What stays with you
- The MCP server and the service accounts
- The policies, approvals and audit trail
- The choice of the model that drives the agent: hosted on your premises or with the provider of your choice
Capabilities involved
- AI agents
MCP server, service accounts, policies and approval.
- Orchestration and compute
Runs, retries, alerts and live logs.
- Governance
Audit of humans and agents, revocation.
- Generative AI
A model served on your premises to drive the agent.
Frequently asked questions
Can the agent delete data?
Not by default: no delete tool is enabled. You decide whether to open one, project by project, and you can require human approval for it.
Which model drives the agent?
The one you choose: a model served on your premises by the LLM gateway, or an external provider. The model only receives what the tools it calls return.
Can we start read-only?
Yes, and it is the recommended starting point: the agent diagnoses and suggests, your teams act.
How do we stop an agent?
By revoking its service account: its next calls are refused. A run it started can be cancelled from the console.
Hand a first on-call duty to an agent
In 8 weeks, a read-only agent, then the actions you choose to open to it.