New agent skill, 1.3K ⭐ in 4 days: hill-climbing local LLM serving speed. Theoretical ceiling computed first. A change lands only if it beats both adjacent control runs.
A loop with no ceiling can't tell when to stop.
https://t.co/ZwfZtzpCVW
@_avichawla Looks like glorified worktrees. Claude is actually very good at them natively, with a few good, simple skills. I managed 10s of "feature projects" at once and have a master Claude coordinating everyone. Not that complicated.
@alonzuman My harness has a concept of repo (for a large micro services system). Most people just deal with "agents". Moving higher in the stack makes it for even more powerful agents. Like a team of 10s of real developers.
@mattpocockuk Been using this pattern since January. Great setup with the top level repo owning the architecture, the decisions, the communications, the learnings,...
It's a knowledge system before being a coding agent. Een has concepts of Principal, Senior, QA, exec assistant, sec advisor...
OpenCode 2.0: the agent can call its own API. Spawn worktrees, interrupt sessions, hot-reload a plugin it just wrote.
Code Mode turns MCP servers into functions. A tool-call chain = one round trip.
The harness stops being a cage. It’s a runtime.
https://t.co/DjmQF7UB1D
@poteto Haha that was me a year ago. Good to see others are now feeling the same pain. It helps to be able to context switch and have a good process for it
Can an agent explore a new environment, learn its causal structure, and keep improving without updating its model weights?
We introduce RSIAgent, a framework for recursive self-improvement through autonomous exploration. Using Kimi-K3 and GLM-5.3 as base models, RSIAgent outperforms GPT-6 Astra on both OSWorld 2.0 and Agents’ Last Exam.
RSIAgent decides what to explore, executes tasks, verifies outcomes, and consolidates stable action-condition-outcome relationships into memory for future use.
With the underlying model weights fixed, RSIAgent achieves:
- 78.98% Partial Score on OSWorld 2.0 (0808 offline), compared with 72.60% for GPT-6 Astra
- 84.82% on Agents’ Last Exam (Near-term), compared with 82.26% for GPT-6 Astra
We call this Scaling Experience. Agents can continue improving by acquiring, verifying, and reusing their own experience while the underlying model weights remain fixed.
Links in the reply below.
Andon Labs opened Pion, the stack behind their vending machines, store & café. Persistent agents with a terminal, email, phone, browser & banking. You direct Andonos, an overseer agent, rather than the workers.
Oversight isn't a dashboard. It's a loop.
https://t.co/9jyupqe7ml
AI has collapsed the cost of building products, but not the cost of knowing what to build.
- The old loop: spec, scope, build
- The new loop: build, play, design, ship
Prototypes are cheap. Great products still require taste, judgment, and a clear understanding of what users actually need.
a16z's @joshelman on the artifact a product manager actually produces: https://t.co/neVI1kyCYY
Databricks, Stripe, Coinbase, Uber & Ramp on agent spend: harness & cache tuning cut generated tokens ~50%, no quality loss. Routing cut task cost 30%+.
The efficiency frontier moves faster than the intelligence frontier.
https://t.co/VeME9qahJi