Herdr lets developers run and supervise multiple AI coding agents simultaneously

·1 views

Herdr is an open-source agent multiplexer that allows developers to run several AI coding agents in parallel within real terminal sessions, each assigned a separate task. Unlike standard terminal multiplexers such as tmux, Herdr tracks each agent's status — whether it is working, blocked, idle, or done — and exposes a CLI and socket API for scripting and automation. Sessions persist even after the terminal closes, and developers can reattach locally, over SSH, or remotely, including from a phone when an agent needs a quick decision. The tool runs locally on any machine with a terminal and requires no account, telemetry, or hosted dashboard. The core productivity gain is parallelism with reduced context-switching: human attention is directed only where a decision is actually needed.

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)

Why Software Systems Sometimes Need the Best Result, Not Just Any Result

A technical explainer published on DEV Community highlights a fundamental distinction in software engineering: finding a specific item quickly versus finding the best item among many candidates. While hash maps excel at direct lookups using known identifiers like user or order IDs, they fall short when a system must compare options and select the highest-priority one. Real-world examples such as ride-sharing apps assigning the nearest driver, hospital triage treating the most critical patient first, and OS schedulers picking the next process all illustrate priority-based selection. The article argues that recognising this pattern — where comparison replaces retrieval — is key to choosing the right data structure in system design. Understanding this shift, the author notes, helps engineers move from simply locating data to intelligently ranking and acting on it.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Silent test flaw hid cross-runtime game logic mismatch for months

A developer building a browser version of a shipped Android game discovered that the same seed and inputs produced entirely different outcomes across the JVM and TeaVM runtimes — a score of 9 versus 6 with one build dying early. The root cause was that java.util.Random's behavior, while documented as a specific linear congruential generator, is not guaranteed to be replicated identically by foreign runtimes like TeaVM. An existing determinism test had been passing throughout, but it was structurally flawed — it ran the engine twice within the same runtime, making it incapable of detecting disagreements between runtimes. The fix took just ten minutes: replacing java.util.Random with a hand-written LCG so both builds execute identical arithmetic. The developer notes that even a golden-file test pinning one runtime's output would have been equally blind, since only a direct cross-runtime diff can expose this category of bug.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Anthropic's Model Context Protocol Becomes the Universal Standard for AI Integration

Anthropic introduced the Model Context Protocol (MCP) in November 2024 as an open standard designed to connect AI models with external tools, databases, and business systems through a single unified interface. Before MCP, each pairing of an AI application and an external tool required a custom-built integration, causing complexity to scale quadratically as both AI agents and business tools multiplied. MCP adopts a client-server architecture where any MCP-compatible AI client can access any MCP server without bespoke engineering on either side, much like a universal connector. Governance of the protocol has since moved to the Linux Foundation's Agentic AI Foundation, and a major updated specification is expected to be finalized in late July 2026. Adoption has grown rapidly, with monthly SDK downloads rising from around 100,000 at launch to roughly 97 million within 18 months, and major platforms including OpenAI, Google, Microsoft, and Amazon have all added MCP support.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Why the AI Agent Runtime Matters as Much as the Model You Choose

A developer experiment using DeepSeek V4 Flash across two different agent runtimes — Codex and Claude Code — produced notably different outcomes on the same long, cross-file coding task. The findings suggest that a model's effectiveness is shaped not just by its raw capability but by the full runtime stack, including protocol, tools, context handling, and how task completion is defined. DeepSeek has publicly acknowledged adapting V4 Flash for Codex and documented Responses API support, signaling a strategic focus on runtime compatibility. The author proposes that effective agent performance equals model potential multiplied by a 'harness realization rate,' which accounts for protocol matching, tool reliability, and recovery quality. The key takeaway is that fair model comparisons require fixing all runtime variables — and results observed without those controls should be treated as runtime-specific observations, not universal benchmarks.

0 comments Read more at DEV Community