Era X · 2026+
The Agentic Era
By C.W. Jameson · Published 19 May 2026 · Last reviewed 19 May 2026
Where we are at the time of this entry. Computer-use APIs, browser-use agents, multi-agent orchestration. The first credible autonomous loops measured in hours rather than minutes. The capability is unevenly distributed — some operators are running production agents that handle real workloads end-to-end; most are still wiring tool-use pipelines that work most of the time.
Computer Use and Operator
Anthropic's Computer Use (October 2024) and OpenAI's Operator (January 2025) are the first commercial computer-use APIs. Reliability lags every other Claude capability; trajectory is steep.
Long autonomous loops
Claude Opus 4.7 with extended thinking can run two-hour coding loops without operator intervention on well-scoped tasks. The same loop without extended thinking stalls within twenty minutes. The capability gap is the reasoning depth, not the raw model size.
Multi-agent
Sub-agent dispatch became a first-class feature in Claude Code in 2025. The pattern is now standard across serious harnesses — a planner agent dispatching specialist agents (search, write, evaluate). The catch is that sub-agent permission inheritance is uneven across harnesses; one of the playbook's hard-won lessons.
Signature models of the era
- Claude Opus 4.x with extended thinking
- GPT-5 with high reasoning effort
- Anthropic Computer Use
- OpenAI Operator
Technical shifts
- Multi-agent dispatch becomes standard
- Autonomous loops measured in hours
- MCP servers proliferate
Market shifts
- Coding-agent revenue continues to compound
- Computer-use enterprise pilots widespread, production deployments selective
Authentication — is the document from this era?
| Tell | Meaning |
|---|---|
| Agent transcript includes sub-agent dispatch with separate context | Claude Code sub-agents, or equivalent multi-agent harness. |
Agents catalogued in this era
- Claude Opus 4.7 — Anthropic's flagship at the time of this entry: extended thinking on, 1M-token context, the most patient of the frontier models.
- Claude Sonnet 4.6 — The workhorse. Faster than Opus, smarter than Haiku, the model most production deployments end up on.
- Claude Haiku 4.5 — The fast one. Used for sub-agents, high-volume classification, and the parts of an agent loop where capability matters less than throughput.
- Llama 4 — Meta's fourth Llama generation. Three sizes, all-MoE, the open-weights default for serious self-hosters.
- Mistral Large 3 — The European frontier model. Codes well, refuses politely, and keeps your data in the EU.
- Qwen 3 — Alibaba's flagship open-weights family. Best non-English capability among open models, strong on coding.
- Computer Use — Claude pointed at a desktop. Mouse, keyboard, screenshots. The first frontier-model computer-use API.
- OpenAI Operator — OpenAI's browser-only autonomous agent. Lives in a sandboxed Chrome, performs tasks, returns results.
- Anthropic SDK / Managed Agents — Building blocks for custom agents — tool use, prompt caching, batch API, extended thinking, memory.
- Gemini Flash 2.0 — Google's workhorse: fast, cheap, multimodal. The Sonnet of the Gemini family.
- Google NotebookLM — AI notebook with source grounding. Generates audio summaries. No hallucination outside uploaded sources.
- OpenAI Swarm — OpenAI's educational multi-agent framework. Minimal, opinionated, not production-hardened.
- Model Context Protocol (MCP) — The open protocol that standardised tool connections for LLM agents. Now an industry standard.
- Vertex AI Agent Builder — Google Cloud's enterprise agent platform. RAG + Gemini + enterprise controls in one product.
Primary sources
- [1] Anthropic: Computer use — 2024-10-22
- [2] OpenAI: Operator — 2025-01-23
From the Almanac shop
The AI Eras — Pocket Field Guide
Ten eras of AI on a single foldable. The Almanac in your pocket.
$19 — Coming soon