Steering Models Through Context

---

Resources

• Context engineering as an operating discipline — Anthropic connects instructions, tools, retrieved information, just-in-time loading, compaction, and subagent isolation. Use this as the lesson’s organizing model: steering means selecting and maintaining useful context, not continually lengthening a prompt. https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents

• Settings have scope and authority — Claude Code’s settings reference shows user, project, local, and managed configuration alongside permissions, environment variables, and MCP controls. It supports a concrete distinction between contextual guidance and enforced access boundaries for agents and connectors. https://code.claude.com/docs/en/settings

• What actually enters a coding session — Claude Code’s operating model explains the agent loop, tools, environment, conversation history, context window, and session continuity. Use it to trace which state persists, which context is compressed, and how interaction modes affect a turn. https://code.claude.com/docs/en/how-claude-code-works

• Profiles and environment configuration — Codex’s advanced configuration documents reusable profiles and more detailed runtime settings. It provides a second harness example for separating a named working profile from the session’s actual messages, tool access, and environment. https://learn.chatgpt.com/docs/config-file/config-advanced

• The Agent Skills specification — The format defines skill metadata, a Markdown instruction body, and optional scripts, references, and assets. Use it to design bounded, reusable skills and linked resource graphs rather than treating every reference as context that must be loaded immediately. https://agentskills.io/specification

• How a harness selects and loads skills — The Agent Skills integration guide addresses discovery, metadata exposure, activation, and loading supporting resources. Pair it with the format specification to distinguish authoring or generating a skill from the runtime behavior that decides when its instructions and scripts become available. https://agentskills.io/client-implementation/adding-skills-support

• The MCP tool contract — The protocol specification defines discovery, input schemas, calls, result content, and tool errors. This is the precise reference for showing how a connector’s capabilities become model-visible and why annotations or hints are not permission enforcement. https://modelcontextprotocol.io/specification/2025-06-18/server/tools

• Writing tools an agent can recover with — Anthropic discusses tool selection, namespacing, descriptions, context-efficient responses, and informative errors. Use it to critique schemas and result payloads, especially whether an agent can recognize failure and choose a useful next action. https://www.anthropic.com/engineering/writing-tools-for-agents

• Code execution as a tool interface — Anthropic describes discovering MCP tools through code and processing intermediate data inside an execution environment. It supports the code-mode versus direct-call comparison, including why Bash or code can reduce context traffic while introducing execution and permission boundaries. https://www.anthropic.com/engineering/code-execution-with-mcp

• Subagents as isolated working contexts — Claude Code documents subagent definitions, tool restrictions, model choices, execution behavior, and persistent memory options. Use it to examine exactly what a delegated agent receives and returns instead of assuming that every agent shares the main conversation. https://code.claude.com/docs/en/sub-agents

• Handoffs are explicit context transfers — The OpenAI Agents SDK documents handoff tools, structured handoff inputs, and filtering the history passed to another agent. This makes a useful contrast with a subagent returning a result: transferring control and sharing context are separate design decisions. https://openai.github.io/openai-agents-python/handoffs/

• Files are not all processed the same way — Gemini’s file-input guide compares ways to supply media and documents to model requests. Use it to reason about text, multimodal attachments, file references, and input constraints rather than assuming that attaching a file guarantees complete understanding. https://ai.google.dev/gemini-api/docs/file-input-methods

• Real-time context is a stream — Gemini’s Live API introduces low-latency audio and video interaction, streaming sessions, and interactive turn handling. It broadens the lesson beyond static threads: incoming media, interruptions, and session state also shape what the model can act on. https://ai.google.dev/gemini-api/docs/live-api

• Lost in the Middle — This paper tests how the position of relevant information affects performance in long contexts. Use it as a counterexample to “it fits, therefore the model will use it,” and as motivation for selection, ordering, summaries, and explicit retrieval. https://arxiv.org/abs/2307.03172