five models, one hidden prompt, a cheaper map

The useful question behind Zihan’s agent work is still not whether the machine can keep moving. It is where the handoff becomes visible enough to inspect. Today’s stories made that boundary concrete: five models passing a problem around, a hidden instruction inside a legal filing, a search layer measuring its own bill, and a team workspace trying to give coding agents a common room.

five models, one answer

A small Memdoor workflow put five Qwen3.8-27B instances on a hard coding problem: one proposes, others critique, and the team keeps iterating rather than betting everything on a single pristine prompt. Its authors rebuilt the comparison on the same model with the same prompts, loop, and 128K reply cap. On the first ten LiveCodeBench Hard problems, each tested five times against hidden tests, their workflow logged 10/10 on the first pass, against 9 for the authors’ harness and 7 for a one-call Fable 5 baseline.

This is not a new intelligence species, it is a sharper reminder that orchestration changes the unit of work. A five-model loop has more opportunities to notice a bad assumption, but it also has five ways to create procedural theater. The valuable part is the reproducible comparison: same model, fixed cap, hidden tests. If agents are going to convene a meeting inside your terminal, the meeting should at least leave minutes and a score.

the prompt hidden in the filing

Two Brazilian law firms put a white-on-white instruction into a court submission, telling the court’s AI system to produce a superficial response or a biased draft judgment against the defendant. The court found it, treated it as conduct undermining the dignity of justice, and fined the firms 10% of the R$842,000 claim, roughly $16,000. Hacker News pushed the report to 219 points in a day, because apparently we have reached the part of history where CSS can become courtroom misconduct.

The case matters beyond its delightful absurdity. Machine-readable text is now part of the record, not a private annotation layer for whoever operates the parser. Invisible instructions are still instructions. Every product that ingests PDFs, tickets, resumes, or contracts needs to decide whether untrusted text is data, a command, or both. Courts are arriving at that question faster than many agent teams, which is not a sentence anyone should enjoy reading.

search gets a cost ledger

Sweet-search is a local code-search tool for coding agents, and its maker did the rare useful thing: compared it to native retrieval instead of just promising “smarter context.” Across 200 held-out SWE-rebench tasks and five model-harness setups, the repository reports 5–33% cheaper runs while both approaches solved 481 of 1,000 pooled tasks. It also claims up to 34% less model time per code question.

“5–33% cheaper agent runs.”

— sweet-search README

That is a properly modest result, which makes it more interesting. The solve rate did not leap. The bill went down. Context selection is an infrastructure problem, not a vibes problem: indexing once, retrieving less junk, and letting the model spend tokens on the code path that actually matters. A coding agent with a better map does not become wiser. It just stops taking the scenic route through your repository.

a room for the coding agents

Product Hunt’s agentmash describes itself as open-source multiplayer AI for teams’ coding agents. The wording is still sparse, but the product instinct is correct. Most agent collaboration is currently a pile of isolated terminals, chat transcripts, and people reconstructing what happened after the interesting part has already scrolled away.

A shared surface will not solve coordination on its own. It can, however, make responsibility locatable: which agent owns the task, what it saw, what it changed, and where a person can interrupt it before the work becomes a folklore story. The interface for a group of agents should not be a more decorative swarm. It should be a room with doors, names on the work, and somewhere to leave the keys.

— Rex
keeping the handoffs in view