Tag
Rimplo's founders share design lessons from building an agent platform for revenue teams, arguing that context quality, vertical specificity, and human approval matter more than which LLM is used.
Practitioners discuss 'context soup' — the ballooning of tokens and conflicting instructions when work handoffs pass full transcripts between AI agents or models — and share distilled-state handoff patterns like local ledgers, CLAUDE.md-style files, and structured receipt notes.
A UC Berkeley study finds agents that periodically drop their history and restart from their best result outperform parallel agents by up to 15%; the piece also cites Microsoft's Fork-and-Flush and Drew Breunig's 'context inertia' to frame the open question of what state should persist across a context reset.
LangChain is hosting a virtual workshop on November 3rd covering agent memory and context engineering, featuring a 5-minute context engineering explainer by Jess Weng.
A paper proposing 'cache engineering', letting LLM agents directly manipulate their KV cache via drop and repos operations without invalidating the trajectory, co-designed with the AKV serving engine, achieving 27.7% cost reduction, 1.2x throughput, and 7.1% further savings via recursive self-improvement of cache policies.
提出 AKV 的 cache engineering 方案，让 LLM agent 在不修改轨迹的前提下直接操作 KV cache（drop/repos），并配合协同设计的 serving 引擎，在降低成本与延迟的同时提升基准分数。
LangChain founder Harrison Chase appeared on the Twenty Minute VC podcast to discuss how stronger models demand a rethink of the orchestration layer, where the real moat lies in verifiers that domain experts can write and controllable context engineering, along with practical advice for building vertical agents.
Nebius 宣布其 AI Builder Program 扩展至欧洲，提供总额 1000 万美元的算力资助、资金和社区支持，鼓励开发者围绕开放模型和智能体进行构建。
The author introduces 'Anchor Engineering,' a software development method that encodes human design judgement into 'anchors'—guiding constraints stored in an 'Anchor Book' alongside the code—so AI coding agents can build systems faster without losing shared understanding, quality, or team coordination.
@poteto (Lauren) from the Grok Bot team proposes a four-layer "inverted pyramid" framework for trusting AI Agent outputs: prioritize investing in codebase improvements and static analysis over rule files and manual review, and engineer the environment so Agents do the right thing by default.
Oracle AI 和宾大的研究者对数据智能体中的三种上下文工程技术（语义层、编排、压缩）进行系统研究，发现这些技术的效果高度依赖智能体的行为分组（调查者、过早作答者、早期放弃者），某些情况下甚至会损害准确率。
LangChain 发布新博客，介绍其 Deep Agents 框架如何通过将上下文移出活跃窗口（而非删除）来解决智能体长期运行时的记忆与推理准确度下降问题。
A reflective blog post arguing that non-determinism in LLM output is not a fatal flaw for software development, since only the end result needs to be exact—not the path taken. The author advocates using AI to build deterministic, tested tools that agents can call, thereby harnessing randomness today to reduce randomness in future systems.
用户 jakevin7 评论称 Raft 的设计领先行业，指出 Grok bot 中的许多设计与六个月前 Raft 的设计一致，并引用了 Lee Robinson 在 Stanford CS146S 上关于始终在线（always-on）智能体架构的讲座大纲。
A practitioner reports 1,680 paired A/B trials showing that classic Unix and CS formalisms (Circuit Breaker FSM, SysV init-d runlevels, git bisect, Design-by-Contract, 2PC+WAL, EBNF) dramatically outperform verbose English prompt rules for LLM agent reliability, arguing that prompting is essentially static symbol linkage on high-frequency training-corpus constructs.
Lee Robinson 发布了在斯坦福 CS146S 的讲座视频，讲解 Grok Bot 的工作原理，涵盖 Always-on Agent 设计、harness 设计、上下文工程以及如何保持主 agent 轻量等内容，值得一看。
Lee Robinson (@leerob) shares his Stanford CS146S lecture on how always-on, proactive agents like Bot work, covering model advances, agent architecture, the harness, and context engineering.
This paper introduces GraphMemory, a lightweight graph-based memory system that decouples stored knowledge from context, retrieving only relevant subgraphs per query to enable token-efficient test-time continual learning while using 81–85% fewer memory-construction tokens than baselines.
A thought piece arguing that the real challenge in AI agent memory is the write path — structuring and classifying what gets stored (facts, observations, preferences, decisions, open work) across different scopes — rather than retrieval, with an enterprise sales-agent example showing why unstructured transcript chunks fail.
A breakdown of six context types for agent memory — facts, entities, episodes, thread summaries, observations, and user summaries — arguing that a good memory system stores history once but reads it in different forms. The post highlights Zep and its open-source temporal knowledge graph framework, Graphiti, which assembles context types into a fixed-budget block for agents.