2024
Reasoning as a separate mode
About this evento1 showed that extra inference-time compute can materially improve hard problem solving.
Capabilities on the main line · PLANNING + TOOLS
Autonomous task execution
In one viewIn 2025–2026, models combined reasoning with search, code, files and application control. They can now receive a multi-step task, not just a question.
This begins the shift from assistant to agent. Today these systems still need human oversight: they accumulate errors, lose the goal on long tasks and learn little from their own experience.
StatusCURRENT
TypeCapabilities on the main line
Marker2026
Events in dossier15
Development chronology
Researched
2024-11
Model Context Protocol
About this eventAnthropic opened a standard for connecting models to data and work tools.
2025-03
An agent platform
About this eventThe Responses API and Agents SDK combined reasoning, tools, orchestration and observability.
2025-03
Autonomous task horizon
About this eventMETR proposed measuring agents by the human task duration they complete at a given reliability.
2025-04
Agent2Agent
About this eventGoogle opened a protocol for agents from different vendors to exchange tasks and results.
2025-10
Designing agent workflows
About this eventAgentKit turned agent applications into an engineering layer with versions, evaluations and traces.
2026-06-09
Claude Fable 5
About this eventAnthropic released a model for long-running coding and knowledge-work tasks designed around hours or days of agentic work rather than a single answer.
Fable 5 matters less as a new peak on short tests than as a change in the unit of work: it was designed for sustained processes involving tools, files and intermediate verification. Anthropic temporarily restricted the model after launch and later restored global access, underlining that longer autonomy requires dependable safeguards as well as capability.
Source 1: Anthropic · запускSource 2: Anthropic · карточка моделиSource 3: Anthropic · повторный запуск2026-06-30
Claude Sonnet 5
About this eventAnthropic updated its mainstream model for coding, agent workflows and everyday knowledge work.
Sonnet 5 shows the other side of agentic progress: long-running workflows must not only be possible on a flagship but fast and affordable enough for routine use. The event therefore belongs on the autonomy line even though the model remains general-purpose rather than a standalone agent.
Source 1: Anthropic · анонсSource 2: Anthropic · документация моделей2026-07
Longer work in a product
About this eventChatGPT Work carries out longer tasks across files, apps and finished deliverables.
2026-09-01
Fable 5.1: sustained projects
About this eventA Claude update for multistep work; evaluations must account for fallback.
2026-09-02
Gemini 3.8 Flash: agentic work
About this eventA GA release expands fast-model capabilities for coding and workflows.
2026-09-03
GPT-6 Astra: computer use
About this eventOpenAI announced a new generation for complex work and application control. Access is phased.
In progress
now
Long-horizon reliability
About this eventThe longer the action chain, the greater the chance of accumulating an error or misunderstanding the goal.
Planned
planned
Memory and learning from outcomes
About this eventAn agent needs to retain experience across tasks and improve without full model retraining.
Distant horizons
ahead
Verifiable autonomy
About this eventA system completes a long task independently while its plan, actions and outcome remain auditable.
Sources and research
Primary material behind this dossier: papers, lab publications and official reports.
OpenAI · 2024Reasoning as a separate modeAnthropic · 2024-11Model Context ProtocolOpenAI · 2025-03An agent platformMETR · 2025-03Autonomous task horizonGoogle · 2025-04Agent2AgentOpenAI · 2025-10Designing agent workflowsAnthropic · запуск · 2026-06-09Claude Fable 5Anthropic · карточка модели · 2026-06-09Claude Fable 5Anthropic · повторный запуск · 2026-06-09Claude Fable 5Anthropic · анонс · 2026-06-30Claude Sonnet 5Anthropic · документация моделей · 2026-06-30Claude Sonnet 5OpenAI · 2026-07Longer work in a productAnthropic · 2026-09-01Fable 5.1: sustained projectsGoogle · 2026-09-02Gemini 3.8 Flash: agentic workOpenAI · 2026-09-03GPT-6 Astra: computer use
Directions
Capabilities on the main line