Releases·Model Release
Dot releases open-source Dot Reflex 14B execution-recovery controller for AI agents
Dot released Dot Reflex 14B, an open-source (Apache-2.0) execution-recovery controller for coding and tool-using AI agents, with weights on Hugging Face. The model is positioned as a controller that recovers failed agent executions, not a replacement for the primary agent. It ships with benchmark evaluation materials in the repo.
97 engagement1 sourceFri, Aug 28, 2026, 02:00 PM
AI Overview
- Weights published on Hugging Face under usedot/Dot-Reflex-14B; repo licensed Apache-2.0 and supports Python 3.10–3.12
- Positioned explicitly as an execution-recovery controller, not a replacement for the agent doing the work
- Repository includes a benchmark overview image under evaluation/, indicating published evaluation results
- Released as a GitHub project (dot-reflex) on 2026-08-28
Entities
Related
Products·building
Agent controller that monitors model dynamics to decide next action
LOLM is an LLM and agent-control system that monitors model dynamics to decide whether to continue, retrieve, verify, branch, or finalize, rather than relying on prompted self-reported confidence. It records whether actions were consumed and produces a run receipt. It helps developers build more reliable multi-step agents by preventing premature finalization.
LOLM is an LLM and agent-control system that monitors model dynamics to decide whether to continue, retrieve, verify, branch, or finalize, rather than relying on prompted self-reported confidence. It records whether actions were consumed and produces a run receipt. It helps developers build more reliable multi-step agents by preventing premature finalization.
1 engagement1 sourceredditFri, Jul 31, 2026, 09:31 PM
Releases·Tool Release
OrcaRouter team releases OrcaReplay, a tool to record, replay offline, and fork AI agent runs across models
The OrcaRouter.ai team released OrcaReplay, an open-source tool for recording, replaying, forking, and debugging AI agent runs with any model. It lets developers reproduce a failed agent run byte-for-byte offline, addressing the pain of debugging agents after the fact. The GitHub launch includes localized documentation in eight languages.
The OrcaRouter.ai team released OrcaReplay, an open-source tool for recording, replaying, forking, and debugging AI agent runs with any model. It lets developers reproduce a failed agent run byte-for-byte offline, addressing the pain of debugging agents after the fact. The GitHub launch includes localized documentation in eight languages.
100 engagement1 sourcegithubSat, Aug 29, 2026, 08:41 AM
Developments·Paper
AgentRewind: Recoverable Execution for Long-Horizon LLM Agents
AgentRewind is a runtime recovery framework that records aligned checkpoints of the agent context and controlled environment state, enabling LLM agents to recover from errors during long-horizon execution. It addresses the problem of error propagation in long-horizon tasks by providing post-error recovery support, which existing methods lack.
AgentRewind is a runtime recovery framework that records aligned checkpoints of the agent context and controlled environment state, enabling LLM agents to recover from errors during long-horizon execution. It addresses the problem of error propagation in long-horizon tasks by providing post-error recovery support, which existing methods lack.
0 engagement1 sourcearxivFri, Aug 14, 2026, 03:20 PM
Developments·Paper
FailFast-RestartSmart: 0.6B monitor predicts SWE agent failures early
Researchers propose FailFast-RestartSmart, a two-stage controller that predicts SWE agent failures early and restarts trajectories to save compute. A lightweight 0.6B monitor detects likely failures, reducing wasted exploration while preserving useful edits.
Researchers propose FailFast-RestartSmart, a two-stage controller that predicts SWE agent failures early and restarts trajectories to save compute. A lightweight 0.6B monitor detects likely failures, reducing wasted exploration while preserving useful edits.
0 engagement1 sourcearxivTue, Aug 4, 2026, 06:55 AM
Developments·Paper
CodeRescue: Budget-Calibrated Recovery Routing for Coding Agents
A new arxiv paper proposes CodeRescue, a method for deciding whether to retry a failed coding agent step with a cheap model or escalate to an expensive one, using execution feedback to optimize cost and success rate. The approach formulates post-failure recovery as a routing problem over a budgeted horizon, potentially improving efficiency in multi-model coding pipelines.
A new arxiv paper proposes CodeRescue, a method for deciding whether to retry a failed coding agent step with a cheap model or escalate to an expensive one, using execution feedback to optimize cost and success rate. The approach formulates post-failure recovery as a routing problem over a budgeted horizon, potentially improving efficiency in multi-model coding pipelines.
0 engagement1 sourcearxivTue, Jul 21, 2026, 05:56 PM