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.
21 engagement1 sourceSat, Oct 3, 2026, 10:52 AM
AI Overview
- A Rust agent-runtime builder keeps retries, context compaction, loop detection, approval gates, and worktree handling in the harness, reasoning in the model
- The builder cites failure cases: repeated tool calls, mid-run command failures, oversized context, rollbacks, restart survival, and premature 'finished' claims
- The builder asks developers of agent runtimes which model-responsibilities they moved to the harness and what should stay model-driven
Related
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.
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 sourceredditSat, Oct 3, 2026, 09:42 AM
Social·Community
What actually makes you trust a local coding agent enough to leave it running unattended?
A Reddit discussion asks what convinces developers to leave local coding agents running unattended for an hour or two. The poster argues trust comes from the harness—permissions, checkpoints, git, tests, rollback, tool restrictions—rather than the model itself, and asks what conditions finally made hands-off operation comfortable.
A Reddit discussion asks what convinces developers to leave local coding agents running unattended for an hour or two. The poster argues trust comes from the harness—permissions, checkpoints, git, tests, rollback, tool restrictions—rather than the model itself, and asks what conditions finally made hands-off operation comfortable.
200 engagement1 sourceredditMon, Sep 14, 2026, 12:05 PM
Social·Community
Developer rebuilds agent harness to fix session memory loss instead of switching models
A developer reports that swapping models did not fix their agent's reliability problems, such as repeating steps or losing memory between sessions. They rebuilt the harness by separating state tracking, context loading, and output checking into distinct components, which resolved the issues.
A developer reports that swapping models did not fix their agent's reliability problems, such as repeating steps or losing memory between sessions. They rebuilt the harness by separating state tracking, context loading, and output checking into distinct components, which resolved the issues.
23 engagement1 sourceredditWed, Aug 19, 2026, 11:57 AM
Releases·Tool Release
Developer builds open-source Toren runtime after agent silently retried failed GPT-5.6 calls for 16 hours
A developer running production agents for customers describes an incident where a model provider rejected an agent's tool calls — GPT-5.6 required a reasoning-effort setting the developer didn't configure — and the runtime retried over 1,200 times across 16 hours with no crash or alert. The incident led them to build Toren, an Apache 2.0 open-source runtime for long-running agents, and take the infrastructure side of agent ops more seriously. The post highlights silent retry loops as an operational failure mode practitioners should guard against.
A developer running production agents for customers describes an incident where a model provider rejected an agent's tool calls — GPT-5.6 required a reasoning-effort setting the developer didn't configure — and the runtime retried over 1,200 times across 16 hours with no crash or alert. The incident led them to build Toren, an Apache 2.0 open-source runtime for long-running agents, and take the infrastructure side of agent ops more seriously. The post highlights silent retry loops as an operational failure mode practitioners should guard against.
14 engagement1 sourceredditTue, Sep 1, 2026, 07:04 PM
Social·Community
Developers discuss budgeting agent retries to distinguish transient failures from token-burning loops
A developer asks how to budget retries in long-running agents without masking real failures, noting that retries differ in kind: timeouts may succeed on retry, weak retrieval needs a new query, and agents can repeat unproductive plans. The discussion centers on moving beyond fixed retry counts to distinguish transient errors from agents wasting tokens.
A developer asks how to budget retries in long-running agents without masking real failures, noting that retries differ in kind: timeouts may succeed on retry, weak retrieval needs a new query, and agents can repeat unproductive plans. The discussion centers on moving beyond fixed retry counts to distinguish transient errors from agents wasting tokens.
8 engagement1 sourceredditWed, Aug 5, 2026, 10:36 PM