17 things your agent harness should have!
Do this first: go down the list and tick what your harness already does. Each unticked item is a failure your agent will hit. The list comes from reading the source of OpenHands SDK , OpenCode , Hermes Agent , DeepSeek harness and Strands harness . Items 1–6 are the basics everyone builds. Items 7–17 are what separates these harnesses from a model in a loop. The rule behind all of them: the…
1. Execution flow: The harness follows a tool loop where the model calls a tool, receives results, and repeats until the model provides an answer. All other features branch off this core loop.
2. Turn summarization: Earlier turns are summarized to allow long-running tasks to continue without excessive repetition.
3. Subagents: The harness can delegate a focused task to a fresh agent with its own set of tools, receiving a single answer back.
4. Skills: Instructions are only loaded when the task requires them, in order to reduce unnecessary tool usage.
5. Model retries: In case of provider errors, the harness backs off and retries the request, defaulting to 6 attempts.
6. Hooks: Custom checks can be run before or after tool calls, on each prompt, and on stop. Hooks can block the call if they exit with code 2.
7. Error messages: The harness provides error messages that name the next step and explain why, instead of just a generic error.
8. Read before edit: The harness refuses to edit a file it has not read, or one that has changed since it was last read, as the edit requires reading the file first.
9. Repeated-call detection: If the same tool is called with the same arguments multiple times, the harness reminds the user at specified intervals and eventually halts the run.
10. Step budget: Once the step budget is spent, the harness sets toolChoice to 'none', appends a wrap-up message, and refuses any further tool calls.
11. End check: The harness ensures the run ends with a proper answer, rather than crashing.
12. Done check: The harness does not trust the "Done" flag; instead, it verifies the run's completion independently.
13. Prompt testing: After a file edit, Hermes sends the model back to run its tests to ensure the edit did not break the system.
14. Judge: Hermes is instructed not to assume the assistant's claim that the goal is met is evidence enough.
15. Tool removal: Instead of blocking tools outright, the harness removes them when they are not needed, reducing the attack surface.
16. Command parsing: OpenCode checks each shell command and its target folder for permission issues before running them.
17. Silent argument repair: Hermes automatically corrects silent arguments, such as converting '42' to '42', before the tool runs, to avoid errors.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.