Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-07-20 · high
The failure sits in how models represent text, not in the prompt, and a two-dimensional view of it mostly closes the gap, per the paper.
Why now: This site covered the paper on July 20, right as agent workflows built on exact text passthrough keep tripping over this same failure.
Frontier models miss on a task simpler than the reasoning they're built for: copying text exactly, per a paper posted to arXiv. It's the bug I hit constantly: agents mangling a long code block, a config file, or a JSON blob on a plain passthrough.
The paper's authors call the failure representational, something no prompt tweak fixes. Feed the model text as a two-dimensional grid instead of a flat token stream, and it copies more accurately. The paper doesn't say by how much.
More context, clearer wording, a stronger system prompt: none of it helps if the problem sits in representation. Telling a model to copy something exactly doesn't fully fix it either, per the paper.
A better prompt won't close this gap. Verifying the output against the source will. Swap the instruction to copy exactly for a diff check on what comes out.
This site covered the paper on July 20, right as agent workflows built on exact text passthrough keep tripping over this same failure.
Each link below shares sources, entities, or timing with this story.
ECP captures agent outputs, tool invocations, and audit context uniformly, with adapters for LangChain, LlamaIndex, CrewAI, and PydanticAI so the same checks run against any of them. arXiv The authors explicitly label it work-in-progress with the method set expected to change....
Astral shipped it July 23 and it hit 251 points and 152 comments on HN (Astral). The default jump surfaces syntax errors and runtime-failure classes previously hidden behind opt-in config. Also new: formatting Python inside Markdown fenced blocks (python/py/python3/py3/pyi/pyc...
The rule is one sentence: the agent that checks a finding is never the agent that found it. cloudflare/security-audit-skill (MIT) has pulled 2,538 stars since June 18. It turns a coding agent into a multi-phase security auditor with a six-phase kill chain: recon, hunt, validat...
PromptResponse ran five semantically identical but syntactically distinct HumanEval variants through GPT-4o over more than 8,200 executions. Consistent formatting, JSON especially, improved generation efficiency and syntactic stability. The LLM-tuned prompts, meaning prompts a...
arXiv 2608.06370 evaluated models emitting code that calls tools against JSON-schema tool calling on BFCL v4. PTC matched or exceeded the baseline in 11 of 14 models, with the GPT-5.6 family up 10.6%, and held stable under parallel execution in 13 of 14. Under context degradat...
tt-a1i/archify is an MIT-licensed agent skill turning a repo or description into architecture, workflow, sequence, data-flow and lifecycle diagrams, exported as self-contained HTML plus PNG, SVG, WebM and a share card. Its verifiability claim is mechanical rather than rhetoric...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.