Approach
Engineers who’ve shipped production agent infrastructure
Last verified: June 2026· about
Fred Brooks taught a generation that adding people to a late project makes it later — the man-month was a myth because humans don’t parallelize cleanly. Agents are different. Tasks that took a team a quarter can be delegated, parallelized, and verified. The shippable agent-month is real — if your infrastructure is ready for it. Most isn’t. That’s the gap we close.
What “the agent-month” means here
The phrase agent-month in the AI world today means two different things. They get conflated, which leads to bad hiring and bad procurement decisions. Here’s the distinction we use:
The mythical agent-month
A Wes McKinney essay (O’Reilly Radar, 2025)
The argument that AI agents are a productivity curve, not a free pass — that adding agents to a late project will make itlater, the way adding engineers did. A critique of naive “AI 10×” claims.
Read Wes McKinney’s essay →The shippable agent-month
This site
The actual deliverable rate of one well-instrumented agent working against a code-base that is ready for it — measured, not estimated. The infrastructure work (LLM cost routing, evals, MCP, agent-readiness) that turns mythical productivity into shipped work.
Read the long-form essay →How we operate
Ship software, not slideware
Every engagement ends with working code you own — routing, evals, MCP servers, dashboards — plus the runbooks to keep it running without us.
Get to numbers fast
We start with a read-only audit and quantify the gap before you commit to a build. Architecture and tradeoffs over buzzwords.
Hire-out is the goal
You should build an internal AI platform team. We help you ship now and hire later, then transition out cleanly. Honest framing wins.
Model- and vendor-agnostic
We default to the most capable models and route for cost where it makes sense. No lock-in, no preferred-vendor kickbacks — just the right tool per task.
Who you work with
Founder & Principal Engineer
Engineer who builds production agent infrastructure — multi-agent coding harnesses, LLM cost routers, MCP servers, and eval platforms. Open-source author behind brat, fast-litellm, and route-switch.
Amrita Sarkar
Marketing
Leads content, growth, and the listicle-aggregator pipeline that gets Agent Month in front of senior engineering buyers before they ever hit the homepage.
Open source track record
The infrastructure we ship for clients is the same infrastructure we build, in the open, under our own name.
Rust acceleration for LiteLLM — faster connection pooling, rate limiting, and memory-intensive workloads.
route-switchIntelligent LLM routingLLM routing with automatic prompt optimization — sending each request to the model that fits.
bratMulti-agent coding harnessA multi-agent harness for AI coding tools — crash-safe state, parallel execution, one CLI.
ormaiSafe agent database accessGives AI agents database access without the risk — scoped, governed access for agents.
agentvfsAgent execution boundaryA workspace runtime and execution boundary for AI agents — guardrails on what an agent can touch.
openclawOSAI assistant infrastructureAn OS-like architecture for AI assistants — a kernel-based design with process-isolated apps.
Work with us
If your team is shipping AI and the costs, quality, or velocity aren’t where they should be — let’s talk.
How we are set up, and what that rules out
Agent Month is the production agent infrastructure practice of Neul Labs Limited, a company registered in Scotland. That structure is worth stating plainly because it explains the shape of what we do and do not take on. We are a small senior team rather than a pyramid, which means the person who scopes your engagement is one of the people who does it, and it also means we cannot staff a twelve-month programme with fifteen people. Engagements that need that shape should go to a firm built for it, and we will say so.
The deliverable is working software in your repository, and you own it. There is no licensed platform underneath, no component that only runs while we are engaged, and nothing that requires a renewal to keep operating. Where our open-source work is useful we use it, and it is public under permissive terms — the same code, not a crippled version with a commercial edition behind it. That is a deliberate constraint on the business model as much as a feature of it.
We are structured to leave. Runbooks are written as the work happens rather than assembled at the end, decisions are recorded with their reasoning, and we push for an internal owner paired with us from the first week even at part of their time. A practice that needs its clients to stay dependent optimises for different things than one that does not, and most of the small decisions inside an engagement follow from which of those you are.
What we will not do: write a strategy deck, put an inexperienced engineer on your work and bill them as senior, assert something about your system we have not verified, or take an engagement we do not think will pay for itself. The last of those costs us deals on calls and it is the single practice that has done most for the quality of the work we do take.