Term
Pi Durable kept two autonomous maintainers working for four days, including through an upgrade of Pi itself. With no real demand, they wrote their own issues and merged 693 pull requests.
CLM reads the situation and each option separately and picks the closest one. It is fast when the options repeat, and weak when the sentence needs nuance.
Coding agents look more capable when every task starts from a clean checkout. New benchmarks show the cost that appears when patches, decisions, and technical debt carry into the next job.
After nearly 10 billion tokens in Codex, why GPT-5.6 Luna became my default model: cost, performance, context, and limits.
The next wave is not just writing code with a copilot. It is coordinating supervised agents, context, tests, and boundaries across the entire development lifecycle.
LLMs can converse, but they do not share a stable domain model. Ontologies, knowledge graphs, and semantic validation can give agents a common vocabulary, safer actions, and verifiable memory.
When an agent is allowed to act too broadly, the risk is not only leakage or abuse. It also becomes hard to operate: it works outside scope, repeats actions, burns budget, and fails without a clean way back.
The agent is no longer the whole product. The next jump is the layer above it: memory across sessions, coordination across agents, context across repos, and automatic optimization of the harness itself.
The generic agent looks simple until it becomes a giant context window with too many tools, too many permissions, and too much cost. Domain-specific agents trade that accumulation for composition, clear boundaries, and cheaper execution.
Your agent nailed the demo and everyone loved it. But how do you know it actually works? If the answer is 'we tested it and it seemed fine', you are operating in vibes mode. And vibes don't scale.