Anthropic — Agentic AI Overview
Overview
Anthropic builds the Claude model family with a focus on safety, interpretability, and reliable long-horizon task performance. Their public engineering writing and Model Context Protocol (MCP) specification are widely referenced reference points for context engineering and agent tool use across the industry.
Key Offerings
| Product / Area | One-liner | Wiki Reference |
|---|---|---|
| Claude Code | Anthropic's agentic coding CLI/IDE — swarm orchestration, KAIROS daemon, 44 feature flags; architecture partially revealed via 2026 source leak | AgenticFrameworks/ai-coding-agents.md |
| Claude Managed Agents | Hosted agent execution platform with persistent memory, dreaming (scheduled memory consolidation), outcomes (self-grading loop), and multiagent orchestration — Harvey reported 6× task completion improvement | AgentPlatforms/claude-managed-agents.md |
| Claude Managed Agents — Dreaming | Scheduled between-session memory curation inspired by hippocampal sleep consolidation; merges patterns, removes outdated facts, surfaces team-wide learnings; research preview | AgentPlatforms/claude-managed-agents.md |
| Claude Managed Agents — Outcomes | Self-grading loop: developer writes plain-language rubric, isolated grader agent scores output, agent iterates; up to +10 pp task success over baseline prompting | AgentPlatforms/claude-managed-agents.md |
| Claude Managed Agents — Multiagent Orchestration | Lead agent delegates to parallel specialist agents, each with own model/prompt/tools, on a shared filesystem with durable event log | AgentPlatforms/claude-managed-agents.md |
| Context Engineering Guidance | Anthropic's practical guidance on managing context windows for long-running agents | ContextEngineering/anthropic.md |
| Context Management API Primitives | First-party API primitives for compaction, tool-result clearing, and cross-session memory — empirically benchmarked on a 328K-token research corpus | ContextEngineering/anthropic.md#context-management-api-primitives-cookbook |
| Model Context Protocol (MCP) | Open standard for tool and resource exposure to LLMs; co-developed by Anthropic | Standards/mcp.md |
| Building Effective Agents | Anthropic's canonical reference on agent design principles and patterns | ProductionBestPractices/README.md |
| Claude Models (Production) | Best practices for using Claude in production agentic workloads | ProductionBestPractices/security.md |
| Agent Skills / SKILLS.md | Reusable, slash-command-triggered workflow definitions for Claude Code agents; SKILLS.md convention packages repeatable engineering procedures as version-controlled, shareable task modules | Standards/skills.md |
| Anthropic Skills Repository | Official GitHub repository of ready-made Claude Code skills — review, deploy, security, docs workflows and more | Standards/skills.md#provider-skills-repositories |
| Anthropic Marketplace Presence | Claude models distributed via AWS Bedrock, Google Vertex AI, and Azure AI Foundry; MCP open ecosystem acts as de-facto tool/agent marketplace | Marketplace/anthropic-marketplace.md |
Sandbox Runtime (srt) |
OS-level process sandboxing for MCP servers — enforces filesystem + network restrictions via macOS Seatbelt / Linux bubblewrap without containers; research preview, Apache-2.0 | SecurityFrameworks/anthropic-sandbox-runtime.md |
| Dynamic Workflows | Claude writes a JavaScript orchestration script; the runtime executes it in background across up to 1,000 subagents — plan lives in code, not in context. Available Pro+ via ultracode keyword or /effort ultracode. Includes /deep-research bundled workflow. |
WorkflowBuilders/dynamic-workflows.md |
| Orchestration Primitives Guide | Decision guide for choosing between MCP, Skills, Subagents, Agent View, Agent Teams, and Dynamic Workflows — includes flowchart, capability matrix, and composition patterns | WorkflowBuilders/claude-orchestration-guide.md |
| Loop Engineering (Claude Code) | /loop, /goal, scheduled tasks, hooks, and GitHub Actions composed with worktrees, skills, MCP connectors, and subagents into self-feeding automation loops |
AgentHarness/loop-engineering.md |
| Agent Client Protocol (ACP) | Claude Code integrates with ACP-compatible editors (e.g., Zed, JetBrains) via an adapter rather than native ACP support | Standards/agent-client-protocol.md |