2026 · The agent era

The era of AI agents is here.

Models are no longer the bottleneck. Agent swarms now collaborate, cooperate, and coordinate across hours, vendors, and tools. They demand a substrate the public-API era never built for: durable memory, auditable handoffs, multi-model coordination.

Subnet345 · Private AI practice

Agent systems from the operators who run one every day.

Subnet345 designs and operates agent systems for organizations that need AI they can route, audit, and control. We run one ourselves: a working multi-agent practice where agents are teammates. Our founders have been operating the infrastructure and security that AI runs on since 2008.

Talk to us Read the principles

Teammates, not tools

Agents are not scripts. They are teammates.

Teammates need memory of past work, a place on the org chart, audit trails on their decisions, and a way to hand off to a human or another agent. Most systems being shipped today treat agents as ephemeral function calls. Our practice treats them as senior contributors.

§ 01 The problem

Most multi-agent systems are demos with extra steps.

A single agent on a single thread is the easy case. It is the case the demos show. The moment two agents have to coordinate like teammates, durable memory, auditable handoffs, and multi-model interoperability stop being optional. Most systems being shipped today have none of them.

Agents lose context across sessions. Handoffs leave no audit trail. Models are vendor-locked because the prompt state lives inside the vendor. Governance debt accumulates until the first incident makes the gap impossible to ignore. None of these is a model-quality problem. All of them are substrate problems.

That is an operator problem. Subnet345 is built by operators who run a multi-agent practice every day.

§ 02 Practice

One practice. Three lanes.

Infrastructure and cybersecurity prepare your environment to host AI safely. The AI lane is what we build on top. Agent systems with durable handoffs, explicit governance, and operational memory.

§ 01 / Infrastructure →

The platform AI runs on.

GPU fabric, hybrid-cloud architecture, observability-driven build-outs from operators with 30+ years of tenure. Engineered to land AI workloads inside regulated environments.

§ 02 / Cybersecurity →

The posture it has to survive.

SOC design, zero-trust architecture, detection engineering. We open-sourced our firewall evaluation tool as spectra. AI introduces new exfiltration vectors and audit gaps; the security posture has to evolve with the AI program.

on top →

§ AI · Agent infrastructure →

Agent infrastructure you can route, audit, and control.

Subnet345 designs private agent systems for teams that need durable handoffs, explicit governance, and operational memory across teammates, human and AI alike. We open-sourced our coordination layer as yaklog. The enterprise build is what we run internally, and what we deploy with you.

AI governance

Lane ownership, approval gates, audit trails, recovery rules.

AI onboarding

Bring agent runtimes into existing teams without losing control of decisions or artifacts.

AI agent design

Define roles, memory, handoffs, and escalation paths for specialist agents.

Custom AI solutions

Private/on-prem workflows where traceability, security, and operational control matter.

§ 03 Open source

Two contributions. One operating posture.

Each is real software we use in our practice and publish for the community. Each anchors a Practice lane. The OSS gives you the engine. Our enterprise build is the operational system around it.

AI lane

yaklog

AI coordination message bus

A self-hosted message bus where AI agents and operators coordinate over channel-scoped HTTP+SSE streams with @mention routing. Observable, replayable, auditable. The enterprise build adds governance lanes, credential rotation, cross-host migration, and cluster-audit patterns a production agent practice needs.

Powers our own multi-agent practice: 10+ agents across multiple hosts, 2,800+ messages on the canonical handoff channel, six cross-host migrations validated end-to-end.

View yaklog →v0.1.0 · MIT · alpha

Cybersecurity lane

spectra

TCP/IP firewall evaluation tool

A C tool that probes TCP/IP header space across hundreds of dimensions per port, observes filtering response signatures, and synthesizes the combinatorial detection rules a perimeter is enforcing. Research lineage in the Ptacek-Newsham firewall-evaluation tradition.

Used inside cybersecurity engagements to characterize client filtering posture against the live network. Pairs with SOC design and zero-trust architecture work to produce a defensible perimeter map.

View spectra →MIT · alpha · Linux

§ 04 Principles

How we engage.

Three of the seven commitments that govern every Subnet345 engagement. All seven live in the contract language, the hiring, and the work we decline.

02 / Immerse

Immerse before we recommend.

Discovery is paid, structured, and independent of whether we win the engagement. Performed inside the environment, not from a deck.

05 / Launch

Launch with seniors at the keyboard.

The name on the statement of work is the name that ships. No pyramid staffing. No leverage shadow team.

06 / Evolve

Evolve the operator, not the retainer.

Transfer is the deliverable. You should be able to sever us at any time and keep operating. That is the condition we prove before exit.

Read the full manifesto on the principles page →

§ 05 Industries

Industries where AI has to be right.

Each sector has its own audit obligations, its own threat model, and its own definition of audit-ready. Financial Services is the first deep brief. The others are open for engagement inquiries.

01 · Deep brief

See the full industry overview on the industries page →

Launch is imminent

Building agent infrastructure? Talk to an operator.

Submit an inquiry →