XNews

Open-source agent projects and agent frameworks

Open-source agent projects and agent frameworks

Open-source agents released MiroThinker, AIGNE and OpenCode Debug. - MiroThinker: open-source search Agent, GitHub ~1.5k stars - Users applied it to Polymarket and stock prediction - Claim: 30B version costs 1/20 of Kimi-K2 (report) - MiroThinker: 235B params, 256K context window - Single-task max 400 tool calls - Benchmarks: HLE-Text 39.2%, GAIA-Val 80.8% - 147k open training samples, SGLang/vLLM deploy - AIGNE paper: 'everything is context' abstraction - AIGNE components: Context Constructor, Loader, Evaluator - AIGNE solves fragmentation and auditing issues - Use cases: memory agents, GitHub assistant via MCP - OpenCode Debug plugin: runtime instrumentation workflow - Debug flow: debug_start, auto-insert probes, reproduce, read logs, clean up - Features: local HTTP server, Ngrok support, .opencode/debug.log - All projects are open-source with public repos

Key insights

MiroThinker emerges as open-source search Agent

MiroThinker is an open-source search-agent product on GitHub with ~1.5k stars; users reported applying it to Polymarket and stock prediction scenarios. The project claims cost-efficiency (a 30B version alleged to cost 1/20th of Kimi-K2), and offers MIT licensing. Reported model specs include 235B parameters, a 256K context window, up to 400 tool calls per task, benchmark metrics HLE-Text 39.2% and GAIA-Val 80.8%, and 147k open training samples with SGLang/vLLM deployment support.

AIGNE: agent file-system abstraction for context

AIGNE proposes 'everything is context' by abstracting AI context as files, implementing a unified agent file-system. Core components are: Context Constructor (mounts and filters relevant files, emits a manifest), Context Loader (reads and assembles content into prompts, supports streaming), and Context Evaluator (validates outputs and writes valuable interaction data back as memory). AIGNE addresses fragmentation, auditability, and maintainability in context engineering and demonstrates use cases like memory-enabled agents and GitHub assistants via an MCP protocol.

OpenCode Debug plugin provides runtime instrumentation

OpenCode Debug Plugin adds Cursor-style debugging for OpenCode: it starts a local HTTP debug server, inserts instrumentation calls (e.g., fetch), reproduces issues, reads runtime data, and cleans inserted code. Features include runtime capture of function inputs/outputs, Ngrok tunneling for remote debugging, persistent logs at .opencode/debug.log, and automated cleanup of instrumented code. The project is open-source with a public repository link.

Open-source tooling is lowering agent development barriers

Collectively, MiroThinker, AIGNE, and OpenCode Debug illustrate an ecosystem trend: open-source agents provide production-capable models and tools, standardized context abstractions, and developer debugging utilities. These projects offer deployable models, context management frameworks, and runtime tooling that together reduce integration and iteration friction for building Agentic applications.

Source posts