Automata from Agent Traces turns traces into automata

// 17d agoRESEARCH PAPER

Automata from Agent Traces turns traces into automata

Researchers compress agent execution traces from 12 datasets into compact finite-state machines with 7–43 states. The structures predict next actions, identify likely failures, and suggest that deployment harnesses shape agent behavior more than the underlying model.

// ANALYSIS

The paper’s strongest idea is treating the harness as observable system behavior rather than invisible glue around an LLM.

–FSMs replay held-out traces at at least 0.997 fitness and build in milliseconds
–FSM context outperforms Agent Workflow Memory for next-step prediction
–Per-state features reach up to 0.94 AUROC for failure prediction
–Runtime monitoring can flag likely failures before agents finish, potentially cutting wasted compute
–The model-agnostic result is promising, but the study does not establish that harness design causally determines behavior

// TAGS

agentsafetyevaluationresearchautomata-from-agent-traces

DISCOVERED

17d ago

2026-08-27

PUBLISHED

17d ago

2026-08-27

RELEVANCE

9/ 10

AUTHOR

omarsar0

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

NEWS1h ago

DeepSeek Code 2.0 leaks reveal 3T model

Leaked specifications suggest DeepSeek is preparing DeepSeek Code 2.0, an open-weight coding model packing over 3 trillion parameters with a 1-million-token context window. The upcoming release reportedly targets native computer-use capabilities to challenge closed frontier models like GPT-6 Astra and Mythos 5.1 in agentic engineering workflows.

NEWS1h ago

OpenAI GPT-6 Sol previews show Astra-level coding

Early leaks reveal OpenAI's upcoming GPT-6 Sol model tier, positioned to deliver near-flagship reasoning and software engineering at significantly reduced inference costs. Benchmark previews show the model executing complex multi-file simulations in single 72,000-token runs to power autonomous developer workflows.

NEWS1h ago

Claude Opus 5.2 spotted internally at Anthropic

Anthropic is reportedly preparing Claude Opus 5.2, an upcoming flagship model inheriting frontier reasoning and coding agent capabilities from the Fable and Mythos generations. Designed to bridge top-tier intelligence with everyday developer workflows, the model aims to deliver high-performance code generation and multi-step reasoning at an accessible price point.

Automata from Agent Traces turns traces into automata