AutoLinx is a network operations platform that adapts to your footprint — from a regional team of 50 devices to nationwide carrier backbones. Agents work against your live device graph, intent rules, and change history. Every decision is engineer-approved. Every action is reversible.
Agents read telemetry directly from devices in real time — BGP state, interfaces, syslog, optical levels. Not aggregated metrics that are already 30 seconds old. The live topology graph builds from this same data → For environments above 25 Gbps aggregate traffic or packet-level diagnosis requirements, AutoLinx pairs with FabricLinx for wire-rate ingest.
Every packet, every BGP UPDATE, every syslog line is read. No 1-in-100, no 1-in-1000. With FabricLinx, the no-sampling guarantee scales to 100G+.
Microbursts, BGP flaps, CRC spikes — agents see them at the moment they happen, not thirty seconds after the fact.
Software-only deployments cover most teams. For carrier-grade throughput or packet-level diagnosis, FabricLinx drops in beneath without changing the agent stack above it.
We don't just send a prompt to an LLM and diff the config. Our agents are grounded in your actual device graph — neighbors, VLAN/VRF topology, intent rules, and change history. The LLM is a tool inside a reasoning pipeline, not the pipeline itself.
Parse config, build a graph of relationships between interfaces, BGP peers, VRFs, ACLs — agents query the graph, not raw text. Root-cause analysis works on relationships, not regex.
Write intent rules your team agrees on — e.g. "core uplinks must be LACP" — and the agent flags drift from intent, not from yesterday's snapshot.
Agents remember who changed what, when, and why — and use that context during root-cause analysis and rollback decisions.
What an AutoLinx agent actually does. These four capabilities aren't separate products — they share the live graph, the audit ledger, the AI stack, and the approval flow. But each has its own behaviors, vendor matrix, and production history worth reading in depth.
A live graph of every device, neighbor, role — built continuously from LLDP, CDP, BGP, SNMP, gNMI. Multi-vendor by design. Updates within seconds of a change.
Describe a change once — Cisco, Juniper, Arista, Huawei, Nokia config emits in parallel. Pre-flight blast radius simulation. Tested rollback every time. Engineer-approved gate.
Live inventory of IP, VLAN, interface, ASN — derived from the discovery graph. Atomic free-resource reservation. Drift detection vs CMDB. The spreadsheet retires.
Continuous policy checks against the live network — SOC 2, ISO 27001, PCI, custom YAML. Drift surfaces in seconds. Evidence packs auto-generate as PDF for auditor review.
In 2022 we made a deliberate choice: wrapping a foundation model wasn't going to work for the APAC enterprise market we serve. GPU procurement is hard. Air-gapped deployments are common. Inference budgets are tight. So we built our own.
A compact transformer-based language model paired with a graph neural network that operates on your live device topology. The LM reasons in natural language; the GNN reasons over structural relationships — interfaces, peers, VRFs, ACLs.
Combined, they outperform frontier LLMs on network reasoning benchmarks while running at a fraction of the compute cost.
Runs on commodity x86 — laptop-class for pilots, single-socket server for production. No GPU procurement, no datacenter density problem, no inference-cost surprise on the cloud bill.
For environments that want more capacity, AutoLinx pairs with FabricLinx (FPGA-based) or augments with frontier LLMs — but the baseline doesn't need them.
Most agentic AI platforms can't deploy where the network actually lives — air-gapped DCs, regulated enterprises, GPU-constrained APAC operators. By building a reasoning stack that runs on CPU, AutoLinx fits into infrastructure that already exists. The buyer who couldn't afford to evaluate frontier-AI-based competitors can evaluate us in a week.
Cisco IOS-XE, NX-OS, Juniper Junos, Arista EOS, MikroTik, Huawei VRP, Nokia SR, Aruba, FRR/SONiC. Agents normalize concepts across vendors so you think in intent, not syntax. Describe the change once; AutoLinx emits vendor-correct config for each. See how intent-to-syntax translation works →
# Same intent — agent emits vendor-correct config intent: core_uplinks_must_be_lacp: match: role=core AND link_type=transport require: "lag.protocol == 'lacp'" on_violation: propose_fix
Agents preserve a chain-of-evidence at every step — config snapshots before and after, telemetry consulted, peers cross-checked, and the reasoning behind every decision. Every action replays. Continuous compliance checks against this same ledger →
Deploys in your VPC. No config leaves your network. LLM calls go through a separate data plane via a redaction layer that strips secrets before they reach inference.
AutoLinx doesn't replace your tools — it works with them. Approvals through Slack, incidents in PagerDuty, changes in Jira/ServiceNow, configs in your Git repo. Your existing dashboards keep working.
Runs in your VPC or on-prem. Config never crosses your network boundary. Air-gap deployments fully supported.
Agents are scoped by role, site, and device class. They can't act outside their boundary, even if instructed to.
Every action lands in an append-only ledger — export for SOC 2, ISO 27001, or PCI on demand.
A 4-week pilot on 100 of your devices, with our solutions team for setup. Measurable outcome by week four — or no commitment.