Tag: ai-agents

A skill's body loads on demand, but its name and description sit in a listing that is always in the prompt, capped at 1% of the context window. Claude Code 2.1.261 adds /skill-doctor to say which loaded skills never get used and what each one costs, so you can prune them before the budget starts evicting the skills you do use.

The cold start of a subagent is not what costs you money. Measured across 117 real Claude Code subagent transcripts, startup context was 0.6% to 18% of each agent's total spend. Depth is expensive because every layer runs its own agentic loop and hands the layer above a summary instead of evidence. Here is the compression test that tells you which layer to delete.

Conversation history can live in your process, in your database, in the provider, or on the machine that ran the agent. Where you put it does not change your token bill, it decides your retention exposure, and it decides whether you can ever change providers. The numbers, the retention windows, and a storage shape that survives all three.

Until now the only way to hand every developer the same MCP servers was managed-mcp.json, a file at a system path that takes exclusive control of MCP. Claude Code 2.1.259 adds a managedMcpServers setting for HTTP and SSE servers, and quietly narrows what allowedMcpServers governs.

Defensive system prompts are heuristic. Information-flow control is not: label every piece of content with integrity and confidentiality, propagate most-restrictive-wins through every tool call, and check the label at the sink before it runs. An 80-line runnable harness, the Agent Framework FIDES implementation, and the false positives the pattern buys you.

ModelContextProtocol 2.0.0 flipped HTTP transport to stateless by default, which is the one change that can drop every client still speaking 2025-11-25. HttpServerSessionMode.StatefulForInitializeClients in 2.2.0 serves both revisions on one endpoint. Here is the ordered upgrade, the ten breaking changes that matter, and the discover-probe bug that only a 2.1.0 client fixes.

Instructions files are deterministic and always sent. AGENTS.md sits at the bottom of the documented repository precedence list. Copilot Memory is a separate store Copilot writes for itself, read by only three surfaces, and it is not in the precedence list at all. Here is the support matrix and the rule for deciding where each convention belongs.

The experimental /rubber-duck command in VS Code 1.135 hands the agent's plan, code, and tests to a model from another family for review. GPT-5.4 critiques Claude, and the cross-family choice is the whole point.

Agent Plugins 1.0.0 standardizes exactly two component types: skills and MCP servers. Slash commands, hooks, subagents, rules, LSP servers, permission config and secrets all stay vendor-specific. Here is the field-by-field matrix across Claude Code, Cursor, Copilot and Codex, and which layout to author in.

GitHub's August 27 and August 28, 2026 changelogs remove the 20,000 line review cap, start reviewing bot-authored PRs, and flip the default review effort from Lite to Balanced on September 28. All three push AI credit consumption up in the same month.