The network-specific agent harness

Agentlang keeps network agents honest.

Every change an agent proposes is scoped, checked against the live network, and accepted only when the network proves it worked.Every agent change is checked against the live network first.

Book a demo →

Runs every ticket Fractl resolves.

Agent proposesAttempt 1

Reset the BGP session from sw-core-02 to dist-02Reset BGP: sw-core-02 → dist-02

  • Austin campus would lose its only route
Blocked by AgentlangNever reached your network
Agent proposesAttempt 2

Root cause: BFD timers of 50 ms × 3 flap the session under load. Set both ends to 300 ms × 3.Root cause: BFD timers too tight. Relax to 300 ms.

  • Austin campus keeps its route
  • Hitless change: no session reset
  • Approved · re-test: no flaps in 30 min
AcceptedConfig change on file, rollback ready

The agent proposes. The harness rules.The harness rules

Four things an agent can't do on your network.

Can't

Act on stale information

Agents reason over the live network map, built from LLDP, SNMP, routing and MAC tables, not over what the CMDB last recorded.

Can't

Reach outside the ticket

Device access and vendor skills load only for what the ticket covers. Everything else is out of reach.

Can't

Call a guess a fix

A fix counts only when the network passes the re-test. Anything else is rolled back and handed to your team.

Can't

Work off the record

Every probe, decision and change is recorded and can be replayed step by step.

Swipe for more →

Autonomy

Autonomy is earned, one workflow at a time.

Every workflow starts supervised. Agentlang moves it up only on evidence, and drops it back the moment a verification fails.

Supervised

Approve every action

Where every workflow starts.

Bounded

Approve risky changes

Proven changes run unattended.

Autonomous

Runs unattended

Promoted after a clean, verified record.

Promotion takes shadow-mode agreement, a zero-harm record and proven rollback.One failed verification drops a workflow back a level. One click pauses everything.

Change control

Every change passes the same five gates.

The model picks the fix. The harness decides whether it runs, runs it, and proves it worked.

Gate 1

Diff + rollback

The exact change and its undo, prepared before anything runs.

Gate 2

Approval

Consequential actions wait for a person.

Gate 3

Live pre-checks

Fails closed if the network changed since the diff.

Gate 4

Execute + re-test

Re-tested against the live network after every change.

Gate 5

Verified, or rolled back

There is no third outcome.

Swipe through the gates →

Built for networks

Rehearsed before it reaches your network.

Proven in a network twin

Every runbook is exercised against real device images with injected faults, and the harness checks network invariants before and after every change.

BGP · OSPFSTP · LAGDNS / DDIVXLAN · IPsecQoSDHCP · NTPNAT · firewall

Speaks your vendors

Vendor skills for the major network families, loaded only for the devices a ticket touches.

CiscoMerakiHPE ArubaAristaJuniperFortinetPalo AltoInfoblox

95% of real tickets resolved at a Fortune 10095% resolved at a Fortune 100

Put the harness on your ticket queue.

Book a demo →