Microsoft Puts AutoGen in Maintenance Mode, Merges It Into Agent Framework

·1 views

Microsoft merged its research-oriented multi-agent framework AutoGen with its enterprise SDK Semantic Kernel into a unified Microsoft Agent Framework in October 2025. AutoGen has been placed in maintenance mode, meaning it will continue to receive security patches but no new features or active development. The two tools had distinct design philosophies — AutoGen prioritized flexible multi-agent conversations while Semantic Kernel focused on enterprise consistency across Python, C#, and Java with tighter Azure integration. Agent Framework aims to combine AutoGen's conversational patterns with Semantic Kernel's production-readiness and governance capabilities. Developers starting new projects are advised to use Agent Framework directly, while those with existing AutoGen systems should plan migration as scheduled technical debt rather than an urgent overhaul.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

LiteLLM disaster recovery requires database, config, and decryption key together

LiteLLM, an AI gateway proxy, is not stateless in team deployments — its PostgreSQL database stores virtual keys, user and team records, budgets, rate limits, and spend history. A full recovery requires three components restored together: the PostgreSQL database, the gateway configuration file, and the decryption key used to protect stored credentials. Restoring only the container or only the database is insufficient to return the system to a working state. The decryption key should be stored separately from the database backup to avoid keeping encrypted data and its key in the same location. Operators are advised to rehearse restores on throwaway instances and treat a never-tested backup as unverified.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Five Structured Output Patterns That Make AI Coding Agents More Reliable

A developer running a fully autonomous coding agent found that parsing AI responses with regex and string matching led to repeated failures in unattended production runs. Ambiguous phrasing, markdown formatting, and mid-answer reasoning changes caused the same parser logic to misread opposite outcomes. The fix was switching from free-form prose responses to schema-validated structured output using forced tool calls via the Anthropic API. Instead of asking the model to describe its decision, the approach forces it to populate a predefined JSON contract, eliminating format ambiguity entirely. The author outlines five specific patterns using Claude's tool-use feature that transformed the agent from unreliable to trustworthy in production.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Solo Developer, 60, Builds AI Memory System by Using It as His Entire Workflow

A self-taught developer with 40 years of experience has spent the past 18 months building and refining a personal AI-assisted productivity system by using it as the backbone of his entire workday, not just as an occasional test. He identified a core problem with AI coding tools like Claude, Cursor, and Copilot: each new session loses prior context, forcing the user to manually re-explain decisions, architecture, and past fixes. To solve this, he built a system where his notes, project decisions, and AI interactions share a persistent memory layer, then continued developing that same system within itself. His real-world daily use — including actual network conditions, ISP quirks, and varied work types such as writing and admin — exposed design flaws no staging environment would have caught. He argues that meaningful dogfooding for solo builders means making your tooling the foundation of how you work, so that when it fails, your whole day fails.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

SWE-Bench Study Shows Top LLMs Resolve Under 2% of Real GitHub Issues

A developer series on DEV Community highlights findings from a Princeton ArXiv study showing that even state-of-the-art large language models struggle to resolve real-world GitHub issues on the SWE-bench benchmark. Claude 2, the top-performing model tested, managed to fix only 1.96% of issues, with all models limited to the simplest problem types. The study reveals that current LLMs break down when tasks require multi-file reasoning, dependency awareness, or systems-level understanding. In response, the author is developing a hybrid multi-persona AI approach targeting medium-complexity issues — the tier where existing models consistently fail. The project uses three specialized reasoning personas covering evaluation, systems analysis, and agentic patch strategies, with further details promised in upcoming posts.

0 comments Read more at DEV Community