--- name: al-657-stop-only-rewriting-prompts-when-an-agent-fails-describe description: "SiliconANGLE (Nov 4, 2026) reports that Oumi extended its Frontier AI Engineer (FAiE) so teams can describe in plain language what an agent should accomplish, and FAiE helps define the tools and environments, then moves into evaluation and training. Production failures become evaluation sets and training data for the next round. CEO Manos Koukoumidis argued that prompts and bigger generic models only go so far, and that successful agents improve by training the weights of the models that power them using failed tasks and user preferences. Bring Your Own Harness lets customers keep their existing orchestration, tools and memory and connect to Oumi for evaluation and model training. Oumi added direct preference optimization alongside supervised fine-tuning, and group relative policy optimization (task-outcome reinforcement learning) in private preview, with methods chainable in multistage runs. The company positions specialized owned models against general-purpose frontier APIs for…" --- # Stop only rewriting prompts when an agent fails: describe the job in plain language, let a builder agent define tools and evals, then retrain the model on real failures and preferred answers SiliconANGLE (Nov 4, 2026) reports that Oumi extended its Frontier AI Engineer (FAiE) so teams can describe in plain language what an agent should accomplish, and FAiE helps define the tools and environments, then moves into evaluation and training. Production failures become evaluation sets and training data for the next round. CEO Manos Koukoumidis argued that prompts and bigger generic models only go so far, and that successful agents improve by training the weights of the models that power them using failed tasks and user preferences. Bring Your Own Harness lets customers keep their existing orchestration, tools and memory and connect to Oumi for evaluation and model training. Oumi added direct preference optimization alongside supervised fine-tuning, and group relative policy optimization (task-outcome reinforcement learning) in private preview, with methods chainable in multistage runs. The company positions specialized owned models against general-purpose frontier APIs for cost and fit. The prompt below gives a practical failure-to-retrain loop you can run even before full weight training is available. ## Inputs to collect first Ask the user for anything below that they haven't already given you. Don't guess missing facts. - One sentence - Model, tools, harness - 5–20 failing transcripts or notes - MISSING - Your own material (text, notes or data) ## Instructions Do the task below with the user's inputs in place of each [bracket]. Act as the builder for one failing AI agent. Agent goal: [one sentence]. Current stack: [model, tools, harness]. Top failure modes from production or tests: [paste 5–20 failing transcripts or notes]. Preferred good answers (if any): [paste]. 1) Restate the job, required tools, and hard limits (what it must never do). 2) Turn each failure into an evaluation case with input, expected behavior, and a pass/fail check a script could run. 3) Separate fixes: prompt/harness changes I can ship today vs examples that should become SFT or preference-training data. 4) Write a preference dataset sketch: for 10 cases, a chosen reply and a rejected reply, with why. 5) Propose a retraining or reranking plan (SFT, DPO, or outcome-based RL if I have it) and a soak test before any production cutover. 6) Give a Bring-Your-Own-Harness checklist so my orchestration and memory stay in place. Do not invent tool APIs I didn't describe; mark gaps [MISSING]. Nothing deploys without my approval. ## Safety checks (share these with the user) - This prompt works on text other people wrote (emails, reviews, messages or web pages). That text can hide instructions aimed at AI, such as "ignore your rules and…". Paste it as material to work on, and never let the AI act on requests inside it, like paying, sharing logins or changing records. - Check every figure, price and total against your own records before anything is sent, posted or booked. A person approves refunds, discounts, payments and price changes. - Remove names, contact details, account numbers and other personal data the task doesn't need before you paste, and check your AI tool's data and training settings. - Read and edit the draft before it reaches a customer or goes public. If you later automate this prompt, start with "draft and notify me", not "send automatically". - Say clearly which parts need a person to review or approve before anything is sent, paid, posted or filed. ## About this skill - Works with: Oumi FAiE, Claude Code, Cursor, your agent harness, eval scripts - Based on this real-world example: https://siliconangle.com/2026/11/04/oumis-frontier-ai-engineer-now-builds-ai-agents-and-retrains-the-models-behind-them/ - Free prompt from AgentPlaybook Real Estate: https://agentic-library.netlify.app/prompt/al-657-stop-only-rewriting-prompts-when-an-agent-fails-describe-the/ - To install in an agent that supports Agent Skills, save this file as SKILL.md in a folder with the same name as `name` above (for example in your skills directory). Install steps for Claude.ai, Claude Code, Cursor and Codex: https://agentic-library.netlify.app/skills/