A Cardano stake pool operated by AI agents: daily checks, safe restart windows and human decisions where keys and money are involved
An independent Cardano stake pool operator.
Reviewed by Christopher Clayton, CEO
- Scale
- 1 block producer, offline cold keys, daily update checks, 1 restart policy, 1 rewards pipeline, 2 kinds of human confirmation
- Engagement
- Continuous operation on mainnet, with agent-run maintenance sessions since October 2026
The challenge
A stake pool earns only when its block producer is online and forging at the moment a slot is assigned to it. The producer must be patched, its hot keys renewed on a fixed cycle and its software upgraded, and every one of those steps can cost a block if it is done at the wrong minute. The operator wanted the routine work done carefully and on time by agents, with a hard stop wherever a mistake would be expensive or impossible to undo.
A Cardano stake pool is a small business that lives or dies on timing. The pool earns a reward only if its block producer is up and forging when the network assigns it a slot, so every restart, patch or key change has a price if it lands at the wrong minute. The operator wanted that routine work handled by custom AI agents that follow the pool's own rules, and wanted every dangerous step to end with a person.
The agents start with a daily check. They look for pending operating-system updates, for services that must be restarted to load patched code and for a reboot that is waiting, and they write what they find to the pool's log. The node itself is protected from automatic restarts, so the agents decide when it is safe. Their restart policy is strict: estimate the outage, multiply it by ten to get the window, read the leadership schedule and go ahead only when no block is assigned in that window. Otherwise they pick the next clear hour and say when.
Keys and money are where the line to a human is drawn. The cold key never leaves an encrypted volume that only a person mounts. An agent can prepare the new hot key and certificate and verify the counter on the chain, but the signing step waits for the person. Software upgrades are proposed with release notes, a checksum-verified package and a rollback copy. The rewards pipeline has its own written rules: it never changes an address by itself, the destination is confirmed by the operator for each run, and after a transient error it stops with its state saved and asks before it resumes.
The result is routine work that gets done on time and leaves a record. A planned reboot with 18 package updates and a new kernel on 26 October paused forging for 2 minutes 18 seconds, hours before the next assigned slot. This is a small, real example of the pattern we build for clients: agents that handle the repetitive, rule-bound work and know exactly when to stop and ask.
What our team delivered
- A daily agent check of the producer: pending operating-system updates, services that need a restart to load patched code and whether a reboot is waiting, with each finding written to the pool's log
- A binding restart policy that the agents follow: estimate the outage, multiply it by ten for the window, read the pool's leadership schedule and restart only when no block is assigned in that window; the node itself is never restarted automatically by the update tools
- Hot-key renewal on a fixed cycle: an agent prepares the new key material and operational certificate and checks the certificate counter against the chain, while the offline cold key stays on an encrypted volume that only a person mounts
- Software releases tracked by an agent that reads the release notes and proposes the upgrade with a checksum-verified package and the old binaries kept for rollback; a person approves
- An epoch rewards pipeline run by agents under its own written rules: addresses and destinations never change on their own, a person confirms the destination in the same conversation for every run, and the runner stops safely and asks before resuming after a transient error
- Escalation by design: every step that touches keys, money or live firewall and login settings is either handled under a written rule or turned into a numbered question for the operator, and each action is logged with what, when and why
Result
Routine upkeep now happens without the operator starting it. On 26 October 2026 an agent-run maintenance session applied 18 package updates and a new kernel with a planned reboot; forging was paused for 2 minutes 18 seconds, in a window the leadership schedule showed was clear, with the next assigned slot hours away. The node upgrade and hot-key renewal on 20 October followed the same pattern, and the daily check has kept pending updates visible instead of silent. When the rewards pipeline met a transient exchange error it stopped, saved its state and asked before resuming, which is the behavior the rules were written to produce.
Get this for your company
Related services with fixed prices and a written scope.
AI agent bot, standard
One bot for a task that crosses two systems or handles varied documents, with exception handling and a review queue for your team.
About 3 to 4 weeks
$5,900 per bot
Managed IT and security (MSP)
One monthly price per team member: help desk, device management, patching, backup monitoring and security basics.
Onboarding starts within 10 business days
$100 per team member, per month
Emergency package: security engineer in 1 hour
Is your company in a cybersecurity crisis? Purchase this package and a security engineer will call you within 1 hour.
Callback within 1 hour, 24 hours a day, 365 days a year
$1,950 per incident
Show 3 more
Initial assessment of public endpoints
What an attacker can see and reach from the internet: your domains, mail, websites, remote access and cloud exposure.
Report within 5 business days of kickoff
$1,000 per assessment
Microsoft Copilot and AI readiness
Before you turn Copilot on, find out what it will be able to see, and fix the permissions first.
Report within 8 business days
$2,400 per tenant
Managed security monitoring
Endpoint detection and response watched by our analysts, priced per device.
Agents deployed within 7 business days
$18 per endpoint, per month
After you purchase, we contact you within 1 hour, any time of day or night, to start onboarding.
No payment is taken online. When we call you within the hour we confirm the scope and send an invoice (ACH, card or crypto).
Critical patches in days, not months
Distribution of days from vendor patch release to deployment for critical-severity CVEs, before and after a managed patching service. The dashed line marks two weeks: under CISA BOD 22-01, US federal agencies must remediate Known Exploited Vulnerabilities within two weeks for CVEs assigned in 2021 or later (six months for older ones). Private organisations are not bound by it, but it is a sensible benchmark. Illustrative figures.
Three approval tiers for AI agents
Every action an agent can take falls into one of three tiers: it runs on its own, a named person approves it first, or a human always does it. The examples are illustrative; each firm draws its own lines during scoping.
Failover in under a second
ISP A fails at 300 ms. BFD (RFC 5880) with a 50 ms interval and multiplier 3 declares the session down 150 ms later, BGP withdraws the routes, and the forwarding table repoints to ISP B. Traffic is back at full rate in well under a second; without BFD the router would wait for the 90 s BGP hold timer. Illustrative figures.
Facing something similar?
Tell us about your environment. We will tell you plainly what we would do and what it would take.