About Kogen Bench

Kogen Bench is the research record behind Kogen, a coding-agent harness built by Almir Sarajčić. It records experiments comparing coding agents, the systems that guide them, and the cost and correctness of their work.

Who is behind it

Kogen is developed by Almir at Optimum Tech, LLC. It is under active development. The goal is to let you shape a software feature and approve its plan, then have Kogen carry it through implementation, checks, independent review, and a resulting commit.

This research site is maintained as part of that work. For inquiries, email [email protected].

The work behind the record

We test model and effort choices, plans, project context, agent execution loops, checks, recovery and implementation languages. Hill-climbing connects those experiments: change part of the process, measure the result, then decide what to test next. Fresh tasks and replication help check whether a promising change holds up.

The homepage introduces these research questions. The round pages distinguish completed experiments from designs, withdrawn work and results still being checked.

Reading the evidence

Start with the findings, then follow the hypotheses to the experiment records. Each round is the authority for its own design, history, outcomes, and limits. Designs are identified as pre-registered or retrospective; descriptive results keep their label.

The site is generated from the public repository. Pages link to the source revision behind the snapshot, and data downloads preserve the distinction between official outcomes and captured deliveries. Missing values and incomplete records remain visible.

Source and reproduction

The repository is licensed under Apache-2.0 with closed contributions: it does not accept issues or pull requests. Kogen brand assets are not covered by that software license.

The reproduction guide explains which public exports can be rebuilt and which executions cannot be replayed. Available task bases, hidden suites and graders are published for inspection and post-run grading. Hidden means kept from the agent during execution. The verification guide and rerun guide state what is available and what remains incomplete. See credits for the upstream projects behind the tasks. Kogen.dev owns product information; this site owns the research presentation.