As agents evolve from simple text generation to executing full tasks and pursuing open-ended goals, scoping goals becomes a critical discipline. The article argues that well-designed goals should make feedback easy to give while the agent operates, addressing failures caused by missing context and enabling continuous learning toward greater autonomy. The author predicts prompt and context engineering will fade while goal engineering endures.
Context Engineering
上下文工程
Engineering practice of dynamically assembling and injecting the right context for models or agents to improve task performance.
Curated Intelligence
Greg Isenberg lays out a practical playbook for becoming the most AI-native company: build a personal second brain in Obsidian, capture voice notes via Wispr and meeting transcripts via Granola, then feed that private context to agents through CLAUDE.md or AGENTS.md instruction files, cross-session memory tools like claude-mem and mem0, and an Airbyte-to-ClickHouse data pipeline. His core argument: models are identical for everyone, so proprietary context is the only defensible moat.
The article reviews how an AI investment-research platform used a coding agent to generate and integrate 37 research skills in four days, then hit three production-grade failures: poor routing for colloquial user queries, context bloat from ~6,000-token per-skill prompts with mutually interfering rules, and untrustworthy outputs where citations were mislabeled and user-specified time ranges were silently relaxed. The author argues none of these stem from business logic but from engineering concerns—routing, structure, rule boundaries and evaluation—and distills a tuning framework around accurate routing, lean structure, strict rules, realistic cases and solid evals.
LangChain detailed its internal GTM Agent, which produced 1,540 sales artifacts in 12 weeks and saved thousands of rep hours. The team attributes most of the quality to context engineering rather than prompt tuning, and reports that deals with a sent artifact advanced 47% of the time versus 16% for matched controls. The approach is reproducible on the Deep Agents harness.
ByteAsk launched an open-source C/C++ coding agent whose upfront context engine locates the fix site before the model reads code, cutting tokens per run from 1,795K to 395K (about 78%) and cost by 68.1% while fixing the same number of bugs. It also flags sibling sites needing the same edit and validates patches only against covering tests, saving roughly 80% on obscure projects and hard issues.
Angel investor Greg Isenberg summarized the current startup playbook in one chart: acquire valuable proprietary context, orchestrate AI agents to build on it, feed the output back into the loop, and call humans like an API. He argues the company with the best context and fastest loop wins. The claim is experiential, with no data or verifiable cases attached.
Anthropic's engineering blog uses human memory as an analogy to break down the components of AI agent memory, framing it under agents and context engineering. It is a conceptual engineering write-up, with the middle of the text omitted and no verifiable methods or parameters provided.
The article argues that collaboration is fundamentally about transferring context rather than exchanging messages, and that context loss compounds multiplicatively once multiple agents join the loop. Using its internal tool Siboot as an example, the author proposes redesigning group-chat collaboration around context flow, isolation and hand-off. It remains a product-philosophy essay without verifiable technical parameters or benchmark data.
AICon Beijing will run Dec 18-19 with 13 tracks spanning agent engineering, inference optimization, and agent evaluation and safety. Tencent engineer Chen Siying will present an AI-native enterprise dev toolchain in the Agentic Software Engineering track, covering tool protocolization, agent runtime engineering, context engineering platforms, and a friction-signal feedback loop. The piece is primarily an agenda announcement without concrete technical parameters or verifiable results.
Deep Insight releases an open-source reference implementation demonstrating a four-layer context engineering strategy to resolve bottlenecks in production AI agents. By combining write-and-execute tools with on-demand skill loading, the approach reduces script token consumption by ~95% and system prompt size by ~97%, enabling efficient context management within limited windows.