Updated
How do I run AI agent fleets in production?
Running agent fleets in production is an operations discipline, not a framework choice. Four problems decide whether it works: reliability (agents fail silently), observability (which step broke?), context (agents reread and drift), and cost (tokens compound across the fleet). Each needs a verified, measured answer, not a bigger prompt. tsukumo runs agent fleets in production to ship its own software, open-sources a tool for each problem, and consults on the operating model.
The four operating problems
Most teams pick a framework and assume production follows. The hard part is operational. Reliability: agents report success they did not achieve, so you gate on a verified state change, not a self-report. Observability: when a fleet fails you need the full execution trace to find which step broke, not output-only logs. Context: agents reread the repo and drift, so you serve one canonical doc per query instead. Cost: token spend compounds across every agent and teammate, so you cut the repeated retrieval once, centrally. tsukumo runs this in production daily and open-sources a tool for each part: yoru for observability, trovex for context, dokan for deterministic runs, wrai.th for coordination. The tools are open; the operating model is what we consult on. Today that is a discipline and a set of open components, not one finished platform.
FAQ
What is agent operations (agent-ops)?
The discipline of running AI agent fleets reliably in production. It covers four operating problems: reliability, observability, context, and cost. These decide whether agents work outside a demo. It is operations work, not a model or framework choice.
What are the four operating problems?
Reliability: agents report success they did not achieve, so you gate on a verified state change. Observability: when a fleet fails you need the full execution trace to find which step broke. Context: agents reread the repo and drift, so you serve one canonical source per query. Cost: token spend compounds across every agent and teammate, so you cut the repeated retrieval centrally.
Do I need a platform to run agents in production?
Not a single product. You need the operating model plus a tool for each problem. tsukumo open-sources a tool for each part (yoru, trovex, dokan, wrai.th) and consults on the model. Today that is a discipline and a set of open components, not one finished platform.
How does tsukumo fit?
tsukumo is a Swiss dev studio that runs agent fleets in production to ship its own software. That operation is the proof. It open-sources a tool for each operating problem and consults with teams on running the model at scale.
Want help running agents in production?
tsukumo runs agent fleets in production to ship its own software, and consults with teams on the operating model: reliability, observability, context, cost. Open tools for each part, a measured approach, no overclaim. The context piece runs locally on your machine:
uv tool install trovex
Open source for each operating problem. tsukumo consults on the model; the tools run on your machine.