|
October 24, 2025
The prompt is dead, long live the spec
|
Prompt-based development reached a limit: inconsistent results, manual tuning, and reward drift. The next phase of AI engineering prioritizes structure over improvisation. Reinforcement learning, clear specifications, and consistent evaluation cycles are redefining the relationship between models and engineers. AI progress now depends less on creativity in prompting and more on clarity in systems design.
|
TL;DR
| • |
Model feedback takes the lead: Reinforcement learning replaces manual prompting at scale.
|
| • |
Reinforcement in production: Practical implementations show where automated feedback loops outperform human labeling.
|
| • |
Specs over prompts: Defined input-output contracts are becoming the dominant workflow for dependable AI development.
|
| • |
[Virtual event] The $2B migration consulting industry is over: See how enterprises are replacing migration consultants with AI agents —live Nov 20, 7:00 PM CET.
|
| • |
Ona changelog: Mobile code review and built-in Mermaid diagram generation.
|
|
This paper formalizes reinforcement learning from model feedback as a scalable alternative to manual prompting. It shows how feedback generated by the model itself can reinforce consistent performance without constant human input. For teams building long-term AI systems, this represents a path toward sustainable improvement rather than one-off optimization.
Zhang's post translates research into applied practice. It identifies clear points where reinforcement learning works and where human supervision remains essential. The balance he proposes helps practitioners design feedback systems that resist reward hacking and maintain data quality across iterations.
This work presents character-level optimization as a precise alternative to token-level prompting. It highlights the gains in efficiency and testability that come from smaller, measurable search spaces. Engineers interested in cost-performance tradeoffs can apply these methods to improve reliability under limited compute budgets.
Fowler's analysis positions specification-driven workflows as a turning point in AI-enabled development. Rather than relying on human intuition, developers define strict input-output contracts. This approach scales AI tooling by emphasizing reproducibility, automation, and the integration of quality assurance into model workflows.
Landgraf connects modern AI practices back to established software engineering principles. His argument is direct: progress in AI will follow the same trajectory as software: version control, testing, modularity, and consistent infrastructure. For organizations seeking maturity, this is a reminder that stability and rigor outperform novelty.
The annual survey confirms what many teams already observe internally: platform engineering is now central to managing AI workflows. The data points show clear trends toward consolidation of tools, governance, and shared standards. Clear evidence that the industry is professionalizing around predictable delivery rather than experimentation. For a deeper look at these trends, read our State of AI Platform Engineering 2025 whitepaper in collaboration with Platform Engineering.
While coding assistants grab headlines, enterprises managing thousands of repositories face a different bottleneck: achieving consistency and maintenance at scale. Agents can now run transformations across hundreds of repos: standardizing, upgrading dependencies, migrating and remediating. This brings a new operational model where centralized teams handle compliance, modernization, and standardization without burdening dev teams.
On Screaming in the Cloud, Ona CTO Chris Weichel discusses the tension of granting agents autonomy to operate independently across hundreds of repositories while maintaining control. And why real productivity gain is from agents working in parallel without constant human oversight, but only if safely orchestrated.
Today, migrations happen manually with spreadsheets, emails and reminders. Agents are now capable of automating CVE remediation, language migrations, and updates at scale. But how do enterprises safely deploy code changes across hundreds of repositories? This session with Platform Engineering explores the infrastructure, governance, and organizational practices to move from manual processes to scaled agent-driven migrations - Join us live on Nov 20th, 7:00 PM CET.
This whitepaper serves as a detailed manual for establishing security baselines across AI workflows. It covers source control boundaries, access segregation between human and agent users, and threat modeling for LLM‑powered automation. It also presents practical policies that reduce exposure when operating across multiple cloud vendors and regulatory regions.
The October changelog introduces three major updates. Developers can now review code changes directly from mobile, with full file diffs and a navigable tree view for modified files. Ona Agent supports Mermaid diagram generation for visualizing architecture, workflows, and dependencies directly in conversations. Ona is also available in Google Cloud, running natively inside your VPC for complete control over security, data residency, and performance.
Platform Engineering Day @ KubeCon + CloudNativeCon NA
Atlanta, Nov 10, 2025
AWS re:Invent
Las Vegas, Dec 1–5, 2025
May your agent diffs be clear and your rollbacks rare,
Your friends at Ona
|