Social·Community
I started building a coding agent around Ollama and ended up caring more about the harness than the model
A developer building Xencode, a Rust terminal coding agent running local models via Ollama and llama.cpp, reports the harness — not the model — became the main engineering challenge. Failure modes like repeated tool calls, bloated context, and unwinding broken sessions pushed logic into the agent runtime. The post invites comparison with others running coding agents on local models.
2 engagement1 sourceSat, Oct 3, 2026, 09:42 AM
AI Overview
- Xencode is a Rust terminal coding agent using Ollama and llama.cpp backends, adding approval checkpoints, rewind, worktree isolation, session state, and planned context handling, retries, and loop detection
Related
Social·Community
Dev abandons LLM coding-agent harness after 8 months and 883 commits, citing overengineering and scope creep
A developer spent 8 months, 883 commits, and billions of tokens building a coding-agent harness that treats the LLM as a replaceable worker, with the system owning planning, state, permissions, verification, and orchestration. The harness verified claimed task completion instead of trusting the model, but kept accumulating features and was ultimately abandoned.
A developer spent 8 months, 883 commits, and billions of tokens building a coding-agent harness that treats the LLM as a replaceable worker, with the system owning planning, state, permissions, verification, and orchestration. The harness verified claimed task completion instead of trusting the model, but kept accumulating features and was ultimately abandoned.
138 engagement1 sourceredditSun, Sep 6, 2026, 12:28 AM
Social·Community
Builder asks where the model-harness boundary should sit in agent runtimes: retries, compaction, rollback, restarts
A developer building an agent runtime in Rust opened a discussion on how to divide responsibility between the model and the harness. The core loop is straightforward, but failure cases — repeated tool calls, partial failures, oversized context, rollback, restarts, and premature completion claims — force reliability logic out of the model. The post asks practitioners which responsibilities they eventually moved into the harness and which should stay model-driven.
A developer building an agent runtime in Rust opened a discussion on how to divide responsibility between the model and the harness. The core loop is straightforward, but failure cases — repeated tool calls, partial failures, oversized context, rollback, restarts, and premature completion claims — force reliability logic out of the model. The post asks practitioners which responsibilities they eventually moved into the harness and which should stay model-driven.
21 engagement1 sourceredditSat, Oct 3, 2026, 10:52 AM
Releases·Tool Release
ollama releases v0.34.1 with MLX memory and cache fixes
Ollama published v0.34.1, a patch release updating its MLX runner and llama.cpp backends. Changes focus on memory handling: the mlxrunner now evicts prefix cache snapshots and waits for evicted runners and free system memory before loading the next model. The release also raises the token repeat limit to 100 and fixes a ChatGPT model selector spacing bug in the app.
Ollama published v0.34.1, a patch release updating its MLX runner and llama.cpp backends. Changes focus on memory handling: the mlxrunner now evicts prefix cache snapshots and waits for evicted runners and free system memory before loading the next model. The release also raises the token repeat limit to 100 and fixes a ChatGPT model selector spacing bug in the app.
14 engagement1 sourcegithubMon, Sep 14, 2026, 10:14 PM
Products·launched
LoopTroop: local GUI to fix context rot in AI coding sessions
LoopTroop is a local, open-source GUI app that acts as an orchestration layer for AI coding agents, addressing context rot during long vibe-coding sessions. It uses an LLM council planning approach and loops to maintain context, helping developers working on complex, long-running tickets.
LoopTroop is a local, open-source GUI app that acts as an orchestration layer for AI coding agents, addressing context rot during long vibe-coding sessions. It uses an LLM council planning approach and loops to maintain context, helping developers working on complex, long-running tickets.
8 engagement2 sourcesredditWed, Aug 5, 2026, 12:51 PM
Social·Community
Users report Ollama local models fail in coding-agent integrations
A developer running Ollama on an M4 Pro Mac with 48 GB unified memory finds direct inference works fine, but the same model fails when used through coding agents. Claude Code with Qwen3.8 completes tasks accurately, while Codex with Qwen3.8 hangs and shows catalog warnings.
A developer running Ollama on an M4 Pro Mac with 48 GB unified memory finds direct inference works fine, but the same model fails when used through coding agents. Claude Code with Qwen3.8 completes tasks accurately, while Codex with Qwen3.8 hangs and shows catalog warnings.
10 engagement2 sourcesredditSun, Aug 16, 2026, 08:05 AM