Introducing Besydr, a program that runs beside your coding agent

Besydr (say: besider) is a program that runs beside a coding agent, such as Claude Code or Codex. It does two jobs. It supervises the work: it checks each request against a checklist, holds the agent to the rules you stated, and narrows oversized tool output to what the task needs. It is also the agent's memory: it turns the corrections in each session into rules, project facts and work memos, and adds the ones that fit the current task to the agent's context.

There is no measured benefit yet. The section "What has been tested" below says what was run.

What goes wrong today

A study of 20,574 coding-agent sessions in 1,639 repositories (Tang et al., EMNLP 2026) labelled 16,118 episodes where the developer had to push back. It covers Cursor, GitHub Copilot, Claude Code, Codex, OpenCode and Gemini CLI. One episode can carry several labels, so the shares add up to more than 100%.

What the developer sees Share of episodes
Agent violates an explicit constraint 38.33%
Agent acts on a wrong reading of the request 26.95%
Agent misreports the status of its own work 22.58%
Code is logically or syntactically wrong 17.82%
Agent misreads the codebase or system state 11.56%
Agent acts beyond the stated scope 10.20%
Commands or tool calls are malformed 2.87%

The study also found that 91.49% of the visible fixes needed the developer to push back, and that constraint violations are 49.49% of episodes in command-line sessions and 32.26% in editor sessions. The table values come from the paper's Table 2. Its appendix was not read for this post.

Besydr is aimed mainly at the first and third rows: an agent that breaks a rule it was given, and an agent that says the work is done when it is not.

What Besydr does

  • Holds the agent to your rules. When you state a rule in a prompt, Besydr repeats it with every later prompt in that session and stores it with your exact words. In later sessions it is added to the agent's context when the task is related. With generated rules switched on, which is off by default, a rule that can be written as a pattern on tool calls may become a guard that denies the matching call and shows your sentence as the reason.
  • Checks the work against a checklist. Before work starts, Besydr writes a checklist of what would show your request is fulfilled, from the request alone. When the agent stops, a model compares the checklist with the tool calls, their results and the diff. If something is missing, the agent continues with the gaps as its next instruction, up to 3 times.
  • Learns from sessions. In the background, Besydr reads a finished session and stores lessons: rules you stated, facts about the project, and a short memo of the work and how it ended. Every rule must quote a message you wrote. A prompt gets at most 2,000 tokens of lessons, chosen for that prompt. A model filters them further, which you can switch off.
  • Narrows long tool output. A plain-text result of more than about 5,000 tokens is cut to the parts the task needs, where the client allows it. The full original stays where the agent can read it.

The agent cannot search the store. Besydr decides what enters the context.

What it does not solve

  • Unclear requests (15.36% of root causes in the study). Besydr does not ask clarifying questions.
  • Wrong code (17.82% of episodes). The completion check catches an unmet request and missing evidence. It is not a code reviewer.
  • Malformed commands (2.87%). Loop detectors notice repeats, not the error itself.
  • Work outside the request (10.20%). A stated rule such as "do not touch unrelated files" is repeated with every request like any rule. A separate check of each call against your intent exists, but it is off by default.

A guard, once generated, covers only what a tool-call pattern can express, and the completion check is only as good as its model. In one test a model passed 65 of 65 runs, 23 of them failures.

What has been tested

In an early A/B test on 6 one-off tasks, turning learning off scored better, turning the completion check off made no difference, and the run-to-run noise was larger than every setting. One-off tasks have no earlier sessions to learn from, so that test cannot measure learning across sessions.

A measurement on repeated sessions in one repository is still owed. Until it exists, this post and the rest of the site claim no result.

Octomind already runs this loop inside its own process. Besydr moves that code out into a separate program so that other agents can use it. The move is checked for parity: octomind with Besydr must behave as octomind did before the move.

Clients and price

Besydr works with octomind, Claude Code, Codex, OpenCode and pi. Octomind gets every feature. The other clients get the features their extension points allow, and the home page lists them one by one. Only octomind has been run with Besydr in live sessions so far. The other four connections are covered by automated tests against each client's documented hook formats.

Besydr is free when used with octomind. With any other client, the local program needs a personal licence after a trial period. A team licence, priced per seat per year, unlocks the team instance and covers the team's use with every client. The team instance is built after the local program and is not available yet. The licence is a file that the program checks offline.

By default, Besydr's model calls go through OctoHub, the model gateway that octomind already uses, with your octomind login. You can set your own provider and key instead. The install page has the details.