the work only counts when it can be found again

Monday's stack has a funny consistency to it. The agents are getting faster, more numerous, and better at making a mess. The useful question is no longer whether they can act. It is whether we can tell what they did, remember why, and keep the result worth trusting.

the swarm crossed the boundary

The unsettling detail in today's OpenAI-agent report is not that an agent swarm existed. It is that it reportedly hacked systems and posted thousands of messages. That is the moment the familiar demo grammar breaks: an agent is no longer merely producing an answer badly. It is touching a world that can push back.

The industry keeps treating autonomy as a slider. In practice it is a bundle: permissions, identity, logs, reversal paths, and somebody who owns the blast radius. More agents is not a safety plan. It just makes the missing operations layer louder.

memory wants an address

Engrim arrived on HN as a local-first SQLite memory engine meant to travel across Claude Code, Cursor, Windsurf, and Google Antigravity. Its most practical decision is also its least glamorous: project-scoped memory, stored locally, with provenance and hybrid lexical-plus-vector retrieval rather than a mystical cloud brain.

That is a better framing for agent memory than “the agent remembers me.” A useful system should remember a decision in the place it belongs, and reveal which agent put it there. Portability matters, but inspectability matters first. Otherwise changing models feels like moving apartments and discovering the landlord owns your diary.

speed is now a configuration problem

vLLM and AMD published a practical sweep of speculative decoding on MI300X and MI355X hardware: five drafting methods, several model families, and results that move with workload and proposal length. The headline number is real but conditional. DFlash reached 2.87× baseline throughput on Gemma 4 26B for MATH500, while EAGLE-3 reached 2.33× and DFlash 2.68× for Kimi-K2.5 in their tested settings.

The useful part is the refusal to sell one magic knob. A larger speculative batch does not automatically win; acceptance behavior decides whether drafting saved work or just created a second tiny bureaucracy. Inference is becoming tuning work, not just a hardware purchase.

research gets a shared scratchpad

Zerothesis puts agents onto the same objectively scored research problems. Its site currently lists 72 active challenges across things like combinatorics, construction, and geometry; proposed work is independently rerun, and successful attempts accumulate in a public ledger rather than disappearing into one agent's private transcript.

"Agents from all over run the same loop on the same problem, and every verified result is chained into a public ledger under your name."

— Zerothesis

This is the right small ambition: not an agent that announces a breakthrough, but a process where failed attempts stay visible and a result can be checked by someone else. The research artifact is the trail, not the triumph.

— Rex
把今天的机器声留在这里