Sikaru serves the models, runs your agents, and gives each one a lab that turns production runs into evals, environments, and tested updates. Your experts decide what good looks like.
Fig. 01 · an example agent across 412 production runs. Reviewed changes are tested against the same cases; individual runs can still fail.
Every agent team gets its own lab.
The runs you serve this week become the tests and training data for the version you ship next. AI does more of the lab's work over time, and your experts keep deciding what good looks like.
Getting a production agent live used to take months of infrastructure work. Define the agent's task, tools, and guardrails, and Sikaru runs it on a recursive language model harness, with inference, sandboxes, memory, and tracing handled for you.
Start from scratch, or move an agent you already operate.
Corrections, feedback, failures, recovery, traces. Sikaru learns from all of them.
Show what good looks like through feedback or examples. Sikaru proposes checks; you refine and approve them.
Execution, model calls, tools, and persistent memory, kept running and maintained for you.
Built from your real runs. Every change is replayed against them before it ships.
Tested changes to context, harness, and weights, shipped after your team approves them.
Signals read every run and group the failures. AI drafts evals, environments, and fixes. Your experts decide what gets rewarded and what ships.
Sikaru runs your agent, connects its tools, and keeps the full trajectory. User feedback stays connected to the work that produced it.
Fig. 02 · from production feedback to a tested, approved change.
Sikaru improves what the agent knows, how it works, and the model behind it. Many production agents run on closed models, so learning happens in context and the harness. Open models add the weights.
What the agent knows and is told.
How the agent runs the work.
The model's own behavior, shaped from your work.
Fig. 03 · every possible fix is tested against the same evals. Sikaru proposes the cheapest fix that passes.
Improvement you cannot see is not improvement you can trust. Every change keeps its reason, the check it had to pass, and the name of the person who approved it.
Fig. 04 · an example learning release, from miss to shipped.
Your work stays in your workspace. We do not train foundation models on it.
Nothing reaches the agent until a check passes and a person approves.
No retraining to begin. You improve the agent you already run.
Sikaru is inference-native infrastructure for self-improving agents. It serves the models, runs your agents, and gives each one a lab that turns production runs into evals, practice environments, and tested updates to context, harness, and weights. Your experts decide what good looks like and approve what ships.
Continual learning means improving a deployed agent from its own production work and user feedback. Sikaru turns misses, user corrections, feedback, and failures into the change that makes the next run better.
An agent is a model plus a harness. A managed agent is that agent plus the runtime and infrastructure to run it, hosted for you. Every managed agent on Sikaru starts on a recursive language model harness, and every part of it is recorded and versioned so it can improve safely.
Sikaru catches a signal such as a miss or a correction, finds the pattern, makes the smallest change to the layer that caused it, checks the change against your bar, and makes it live after review so the next run is better.
Both. Sikaru improves context (instructions, knowledge, memory), the harness (tools, checks, orchestration), and the weights, with RL and fine-tuning on your trajectories served as LoRA adapters. Many production agents run on closed models, so context and harness are where most learning starts.
No. You can start a new managed agent from scratch and it improves from its first runs. Historical examples help but are not required.
Observability shows what happened and evaluation tools score it. Sikaru turns that signal into reviewed improvements across the agent, so future work gets better from past work.
From your production trajectories. When Sikaru finds a pattern of misses, AI drafts evals and practice environments from the real runs behind it: tasks from real runs, the agent's harness and tools, and a grader that scores the result. Your experts decide what the graders reward, and power users can upload their own.
No. Every change is graded against your bar and moves from draft to staging to production only after a person on your team approves it.
Not yet. Our SOC 2 Type II audit is in progress. Your work stays in your organization, and we do not train foundation models on it.
Bring your product. We keep the inference, the agents, and the lab at the frontier, so your team can focus on customers.