01
Cloud incident: GCP logs stop reaching Sentinel
Changefy traces a broken log pipeline back to an IAM rotation, drafts the fix, and confirms events are flowing again.
sink=log-forwarder · status=VERIFIED · lag=42s
CHANGEFY · GOVERNED AI OPERATIONS
Your AI operations engineer.
Tell Changefy what you need. It investigates your environment, builds a plan, waits for a human to authorize it, executes the work, and proves the result — with a full record of every step.
- Policy enforced
- Human authorized
- Evidence captured
- Infrastructure agnostic
Built to operate across your infrastructure
FROM INTENT TO VERIFIED CHANGE
Seven steps between a request and a proven outcome.
- Investigate
- Diagnose
- Plan
- Review
- Execute
- Verify
- Audit
Traditional AI says
“I’ve restarted the log forwarder — it should be working now.”
Changefy did
Re-granted the IAM role, redeployed the sink, restarted the forwarder, and confirmed a live event reached Sentinel 42 seconds later. Recorded in evidence req_9f23.
01 · INVESTIGATE
Understand the estate before touching it.
Changefy pulls real context from the systems involved — not just the request text — before it forms an opinion about what’s wrong.
- Infrastructure context
- Configuration state
- Logs
- Metrics
- Dependencies
- Recent changes
- Policies
- Ownership
02 · DIAGNOSE
Say what’s wrong, and how sure you are.
For the payments p99 regression, Changefy correlated deploy 2c1f4a with the latency spike and identified a connection-pool sizing regression as the root cause, at 92% confidence — stated as a number, not a guess dressed up as certainty.
03 · PLAN
Every plan carries its own risk profile.
Changefy doesn’t just propose actions — it scores them. Risk level, blast radius, rollback path, and the exact checks that will prove success, all before a human ever sees an approval request.
If reality drifts from the plan mid-execution, Changefy stops and asks — it doesn’t improvise past what was authorized.
04 · REVIEW
AI proposes.
Humans authorize.
- Every production-risk plan requires an explicit approval.
- Approvers see the full plan, not a summary.
- Requests can be approved, sent back for changes, or rejected.
- Approval policy is configurable per environment and per risk level.
05 · EXECUTE
Watch it happen, step by step.
Execution runs the authorized plan one step at a time against real systems, streaming a live log — with an abort control always visible.
06 · VERIFY
Don’t trust the AI because it says it worked.
Make it prove it.
07 · AUDIT
Every AI action. Accounted for.
USE CASES
Real operational work, end to end.
02
Resize an undersized production VM
Investigates CPU pressure, proposes a resize window, executes with a rollback snapshot, and verifies headroom afterward.
vm=api-worker-03 · size=D4s→D8s · risk=LOW
03
Tighten an overly permissive firewall rule
Finds a rule open to 0.0.0.0/0, plans a scoped replacement, and verifies no dependent service loses connectivity.
rule=allow-any-8443 · blast_radius=2 services
04
Ship a GitHub change to Azure
Picks up a merged PR, plans the deployment, waits for approval, executes the release, and verifies health checks pass.
pr=#4821 · env=production · rollout=canary
05
Expand a SAN LUN before it fills
Diagnoses disk growth trend, plans a capacity expansion with a maintenance window, and verifies free space afterward.
lun=vol-9c2 · free=4% → 38%
06
Chase down payments p99 latency
Correlates a deploy with a latency regression, proposes a connection-pool fix, and verifies p99 returns to baseline.
deploy=2c1f4a · p99=840ms → 210ms
07
Contain a Sentinel security alert
Investigates a suspicious sign-in alert, plans a scoped containment action, and verifies the session is terminated.
alert=SEN-8842 · action=revoke session
08
Run a staged patching cycle across a host fleet
Patches dev a week ahead of production, drains sessions in capacity-preserving batches, and verifies every host healthy before closing the window — planned maintenance, not firefighting.
hosts=52 · batches=2 · min_capacity=50%
CONNECTIONS
Plugs into the estate you already run.
Cloud
- Azure — Available
- AWS — Available
- Google Cloud — Beta
Observability
- Datadog — Available
- Sentinel — Available
- Grafana — Beta
- Prometheus — Beta
Source control
- GitHub — Available
- GitLab — Planned
Infrastructure
- Terraform — Available
- Kubernetes — Beta
- VMware — Planned
IT / Ops
- ServiceNow — Beta
- Jira — Planned
ARCHITECTURE
Autonomous where it should be.
Constrained where it must be.
- User request
- Reasoning
- Context layer
- Policy engine
- Plan
- Human authorization
- Execution runner
- Target systems
- Verification
- Evidence ledger
- Least privilege on every credential Changefy holds
- Scoped, short-lived credentials per execution
- Policy enforcement before a single action runs
- Human approval gates on anything above low risk
- Hard execution boundaries per environment
- Immutable activity history for every work item
- Secrets stay inside the execution environment
- Stop-on-deviation the moment reality departs from the plan
WHY CHANGEFY
A copilot suggests. An agent might act. Changefy proves it.
| Capability | Traditional Copilot | AI Agent | Changefy |
|---|---|---|---|
| Understands the request | Yes | Yes | Yes |
| Investigates real infrastructure | No | Partial | Yes |
| States a diagnosis with confidence | No | Sometimes | Yes |
| Produces a risk-scored plan | No | Rarely | Yes |
| Requires human authorization | N/A | Optional | Always |
| Executes against real systems | No | Sometimes | Yes |
| Stops on deviation mid-execution | No | No | Yes |
| Verifies the outcome independently | No | No | Yes |
| Produces an exportable evidence trail | No | No | Yes |
INSIDE CHANGEFY
CHANGEFY
Tell Changefy what needs to happen.
Let AI do the operational work.
Investigate. Plan. Authorize. Execute. Verify.