01
Agentic software
Agents that work through tickets, inboxes, CRMs and repositories, do the multi-step work and stop for review before anything is sent, spent or deleted.
AI agent and generative AI development
We build and run your AI agents and generative AI applications
Hand us the project. We design it, build it in your repository, connect it to your systems and run it in production, or hand it to your team. You keep the code, the data, the keys and the final say.
What we build for you
Agents, generative AI apps and the integrations that make them useful
Most AI projects stall between the demo and production, usually on the integration and operations work. We take all of it on, so there is one team to hold to account.
- 02
Generative AI applications
Knowledge assistants that cite your documents, drafting tools and copilots inside the apps your teams already use, running where your data rules allow.
- 03
Integrations with your systems
APIs, MCP servers and connectors into your CRM, ERP, ticketing, databases and file stores, with secrets, logging and access control from the first release.
Benefits of TenthMatrix AI agents
What you get when our agents do the work
Each benefit maps to software we build and run ourselves, so you can inspect it before you sign anything.
A person approves what matters
Agents draft and prepare. Anything that sends, spends or deletes waits for a named person.
See JobshoutSelf-hosted or managed
Run the agents in your own cloud or data centre, or have us run them. The code is the same either way.
Engagement modelsYour data and keys stay with you
Data stays in your systems and model keys stay in your secrets store. Pick the model provider, or host an open-weight model yourself.
Security and IPSecrets isolated per agent
WSL Vault issues short-lived, scoped credentials, encrypted per tenant and recorded in a tamper-evident audit log.
See WSL VaultReleases gated on health
Ring Promoter moves each change from int to prod only when health checks pass, and rolls back when they fail.
See Ring PromoterSecurity at the edge
WSL Proxy puts an API gateway, WAF rules and rate limits in front of the agent and the tools it calls.
See WSL ProxyCost and behaviour you can see
A trace for every run, cost and latency per agent, and alerts when either drifts.
How we measureBuilt by a team that runs agents
Our own agents run in production on this stack. We publish how they work, including what broke.
Read the articles
How outsourcing works
Discovery, pilot, build, then run or hand over
Each stage ends with something you keep, even if you stop there.
01 / 04
Discovery
We pick one job, the systems it touches and who checks the result, and write down what the agent must never do. You get a written scope and a fixed quote for the pilot.
02 / 04
Pilot
The smallest version that does the real job on real data, behind approval gates, measured against the scope. Then a go, change or stop recommendation.
03 / 04
Build
Production code, integrations, secrets, guardrails, tests and a release path, all in your repository. You can read every prompt, trace and test.
04 / 04
Run or hand over
We operate it with monitoring, cost reports and a regular review, or hand it to your team with runbooks and training.
Engagement models
Four ways to work with us
Choose by how clear the job is and how much you want to own at the end.
Fixed-scope project
One agent, application or integration, with a written scope and an agreed price for it.
Dedicated agent team
A small team working through your backlog of AI work, month by month.
Embedded engineers
Our engineers join your team, work in your repos and pass on what they know.
Build, operate, transfer
We build it, run it for an agreed period, then hand it over with runbooks and training.
Proof
Software we build and run ourselves
Client agents run on the same products we use in production. Each one started as something our own agents needed.
Ring Promoter
CI/CD control plane
Promotes every release through int, test, acc and prod rings. Health gates decide, and failed checks roll back on their own.
WSL Proxy
API gateway · CDN · WAF
Sits in front of your origins with live configuration, an edge cache and WAF rules. An MCP gateway for agents is on the roadmap.
- Open source
WSL Vault
Secrets manager
Envelope encryption with a key hierarchy per tenant, dynamic leases and a tamper-evident audit log. Steal the server, not the secrets.
- Open source
Jobshout
Agent orchestration
Specialist agents that work in Gmail, GitHub and live sites, stream every step and stop for your approval before anything goes out.
- Research
Jev
Decision-layer research
Our independent series and benchmark lab on fast decision models, agent harnesses, routing and the cost of inference.
Nebula family
NebulaDB · NebulaDNS · NebulaCR · NebulaCB
Rust infrastructure for agent workloads, from NebulaDB, a database with vector search and streaming RAG, to NebulaDNS, authoritative DNS with Prometheus metrics.
Writing
Latest from the team
Engineering notes, long-form technical articles and the Jev insights series.
Big Prometheus at Scale: Pros, Cons, and a MoSCoW Map of Observability Tools
TenthMatrix rewrite of the Big Prometheus problem space: Thanos, Cortex/Mimir, VictoriaMetrics, M3/Chronosphere — plus MoSCoW across OTEL, Grafana, SaaS APM and more
Observability Is Not Optional: Make It a Business Standard
Why fragmented monitoring creates blind spots, duplicate spend, and slow incidents — and how to make centralised observability a non-negotiable company standard.
WSL Vault Technical Deep Dive: Zero-Trust Secrets, Dual Control, MFA & Multi-Region HA
Technical deep dive: envelope hierarchy, dual-control workflows, MFA, leases, active/active multi-region HA, feature tours, and presentation chapters
Need the hardware?
GPU workstations and clusters live at Workstation
Our sister company sizes, builds and supports the machines that run private models. Tell them what you want to run and they will spec it.
Start a project
Tell us the job. We'll scope the build.
Send the task, the systems it touches and who checks the result. Within one working day an engineer replies with what we would build, what it should not do and how we would prove it.