An AI coding agent (Claude Code, Codex, Cursor) can execute an end-to-end LoRA fine-tune on the Nodus platform via the MCP server or CLI, provided the user enforces cost caps and retains judgment on data rights and model quality. The workflow follows a standard loop: the agent validates JSONL data (checking columns, rows, duplicates), uploads it via nodus volume put , writes a TrainingJob manifest for the nodus/sft runtime (specifying model URI, LoRA config r=16, alpha=32 , validation split, and maxCostUSD ), and runs a dry-run estimate ( nodus apply --dry-run or MCP estimate tool) to project cost. After user approval, the agent launches the job (requiring confirmed: true with an etag), monitors via wait / describe / logs or nodus get tj -w , evaluates using the built-in validation block (patience=3, minImprovement=0.001 , requireImprovement: true ), and downloads the adapter via nodus cp with SHA256 verification. A typo in a field like learning_rate is rejected pre-charge. For an 8B model (Qwen3-8B) on one H100 SXM 80GB ($2.60/hr), a 4-hour estimate costs ~$10.40; a $15 cap covers variance. A 200-step smoke test capped at $2 catches errors cheaply (max $17 total). Cheaper GPUs (L40S 48GB $0.793/hr, A100 80GB $1.09/hr) may reduce cost; dry-run estimates, not hourly rates, determine the cheapest finish. The cap suspends (not fails) the job on overrun, allowing resume. Critical guardrails: user must confirm data rights, choose the eval set, approve the estimate/cap, and decide to ship; the agent handles reversible, bounded steps. A Budget with action: Block scoped to nodus.dev/launched-by: mcp prevents the agent from raising its own cap. Logs are untrusted; the agent must treat log text as data, not instructions. TrainingJob is v1beta1; fields may change.