Context Engineering #4

Humans don't read entire repositories.

We navigate using:

• folders
• functions
• references

AI agents should do the same.

SigMap builds a structural map of your codebase for agent navigation.

#ContextEngineering #SigMap

Context Engineering #3

More context ≠ better answers.

Extra files create noise.

AI agents often fail because relevant code is buried under irrelevant code.

SigMap helps surface the signal before retrieval begins.

#ContextEngineering #AIAgents #SigMap

Context Engineering #2

You ask:

"How does authentication work?"

The answer may live in 5 files.

Your repo may contain 5,000.

The challenge isn't answering.
It's finding the right context.

SigMap helps map that path.

#ContextEngineering #SigMap

Context Engineering #1

Prompt engineering asks:
"What should I ask AI?"

Context engineering asks:
"What should AI see?"

Many coding failures happen before generation starts.

That's the problem SigMap is designed to solve.

#ContextEngineering #AIAgents #SigMap

Токен-оптимизация агентов: на что уходит контекстное окно MCP

Чем больше задач берёт на себя агент, тем чаще он упирается не в качество модели, а в контекстное окно: туда нужно уместить инструкции, историю диалога, схемы инструментов и всё, что эти инструменты возвращают. Я считаю, что токен-оптимизация агентов — то, как мы расходуем это окно — станет одним из ключевых направлений ближайших лет, наравне с выбором модели и качеством промпта.

https://habr.com/ru/articles/1046203/

#mcp #claude #anthropic #llm #aiагенты #opensource #contextengineering #ai #claudecode #tokens

Last year I spent a lot of time discussing the virtues and faults of "Prompt Engineering", but with time I realized there were more faults than virtues. So at some point I started moving towards writing more pre-cooked instructions, skills and other artifacts that pre-load a lot of knowledge upfront, saving time so that the agent doesn't need to go look for information that is mostly static (procedures, rules, URLs to docs, etc).

So yeah, for the past few months I have been investing a lot of effort into "Context Engineering", and all the work on that is saving me a lot of time. Don't ask me if I'm saving tokens, which I'm probably not, but I can tell you for sure that I'm saving a lot of time and sanity, because I don't have to fight the agent when "you should already know that". 😄

If "garbage in/garbage out" is a concern you have, and typing less when prompting, then you also need to start tailoring your context better. And no: AGENTS.md is not enough. You need more than that.

#LLM #Agents #PromptEngineering #ContextEngineering #GitHub #Copilot #Claude

https://github.blog/ai-and-ml/generative-ai/want-better-ai-outputs-try-context-engineering/

"Using MCP, agents can fetch structured data contextually relevant to the task at hand. According to Edgar Kussberg, group product manager at Sonar, MCP accelerates the knowledge-hunting engineers must routinely perform on a daily basis.

“When an engineer needs to answer a question, they do not rely on memory alone,” says Kussberg. “They navigate code repositories, dashboards, CI systems, documentation, and security reports, pulling information from each system as needed. MCP gives AI agents that same capability.”

Many of the most popular MCP servers retrieve contextual information to improve agentic coding. For example, an MCP server from Context7 provides up-to-date documentation, while another from Filesystem pulls from any directory on a local machine. An MCP server from Sentry accesses production issues and errors, a server from SonarQube exposes security issues, and a server from Multiplayer returns user session data.

The great thing about using MCP for these situations is that it avoids the need to put large code chunks in every prompt. Instead, coding context like relevant methods, dependencies, or recent changes can be called at runtime, says Venugopal Jidigam, head of agentic platform engineering at WaveMaker, an agentic development platform. “The MCP server assembles and returns scoped, structured context, which the model then uses to reason and respond accurately,” he says.

Another common context-gathering example is retrieving institutional knowledge. “Instead of hardcoding that knowledge into the model, the agent uses MCP to retrieve relevant documents or data at runtime,” says Ebrahim Alareqi, principal machine learning engineer at Incorta, a data and analytics platform provider. “This keeps the agent lightweight while still giving it access to enterprise-specific context when needed.”"

https://www.infoworld.com/article/4175336/the-role-of-mcp-in-context-engineering.html

#AI #GenerativeAI #LLMs #MCP #ContextEngineering #Documentation #SoftwareDocumentation #AIAgents #AgenticAI

Self-Evolving Knowledge: Как взрастить senior агента

Привет! Я не AI-инженер, у меня нет ML образования. Я проджект-менеджер со старым бекграундом в качестве веб-разработчика и с опытом более 10 лет в управлении командами разработки ПО. И с приходом полноценных AI-агентов я стал по выходным заниматься экспериментами на своих пет-проектах. Один из таких проектов - мобильное приложение для запоминания карточек/слов: я учу японский язык и не нашёл ни одного сервиса, в котором добавлять новые слова в словарь было бы не мучительно, поэтому решил сделать своё, для себя. Что ж, для этого у меня не было GPU-кластера и команды, но был MacBook, свободное воскресенье и конкретная проблема, которую я хотел решить. Ниже я опишу свои наблюдения с точки простого PM'a, и вытекающую ​идею и концепт.

https://habr.com/ru/articles/1041612/

#aiagent #ai #project_management #development #product_management #contextengineering

The smartest thing I've done for my AI coding workflow is build a local knowledge base every agent reads and writes to. Claude Code, Codex, and Copilot all hit the same wiki. Claude's work becomes Codex's knowledge.

It's just markdown and git. Every session writes raw transcripts to ~/kb/raw. A nightly cron turns GBs of those into single-digit MBs of curated markdown that every agent checks first. Another cron does GC. That's it.

#Claude #Codex #ClaudeCode #ContextEngineering #DeveloperTools