Facts · The field · 7 cards

Agents & MCP

Models that use tools, and the plugs that connect them

How long a task AI agents can finish

  1. 2019GPT-2: tasks that take a human expert seconds.
  2. Early 2025Claude 3.7 Sonnet: about an hour.
  3. Aug 2025GPT-5: about 2 h 17 min.
  4. 2027 to 2029 (projection)If the trend held: week-long tasks. METR offers this with caution.
The task length frontier models finish half the time doubled roughly every seven months from 2019 to early 2025. Source: METR, 2025

These facts come up in the course Agents at Work, where each one is explained step by step.

1

~7 months

How often the tasks agents can finish double in length

METR times software tasks by how long a human expert takes. The length frontier models finish half the time doubled roughly every seven months from 2019 to early 2025. METR's January 2026 update found it faster since 2023: about every four months.

How we got this number

Seconds for GPT-2 in 2019, about an hour for Claude 3.7 Sonnet in early 2025, about 2 h 17 min for GPT-5 in August 2025 (METR's original task suite). Extended with caution: week-long tasks around 2027 to 2029.

Source: METR, 2025

2

Using a real desktop by screenshot, mouse and keyboard

Claude 3.5 Sonnet, Oct 2024, OSWorld14.9 % of OSWorld tasks
Claude Sonnet 4.5, Sep 2025, OSWorld-Verified61.4 % of OSWorld tasks

OSWorld is a benchmark of chores in real desktop apps; OSWorld-Verified is a corrected version of it. Humans score about 72 percent.

Source: Anthropic, 2025

3

$1 B

Claude Code run-rate revenue, six months after launch

On 3 December 2025 Anthropic said its coding agent had reached a $1 billion run rate in November. Coding pays first because tests grade the work: pass or fail.

How we got this number

Run rate means one month of revenue multiplied by 12. Claude Code became generally available in May 2025; the figure came about six months later.

Source: Anthropic, 2025

4

Wait, really?

Anthropic published MCP, a standard plug between AI models and tools, in November 2024. Who adopted it?

Show the answer

Its rivals: OpenAI, Google DeepMind and Microsoft.

MCP (Model Context Protocol) lets a tool be wrapped once and used by any app that speaks it. Within a year the competition had adopted it.

Source: Anthropic, 2024

5

8 days

Time for OpenAI, Anthropic and Mistral to add MCP to APIs

In May 2025 three labs shipped API-level MCP support within eight days of one another. In December MCP was handed to the new Agentic AI Foundation, part of the Linux Foundation.

Source: Simon Willison, 2025

6

No. 1

Prompt injection tops the OWASP list of LLM app risks

A web page or tool result can contain instructions, and the model cannot reliably tell them from yours. So careful systems keep risky actions behind human approval.

Source: OWASP Top 10 for LLM Applications, 2025

7

Guess first

A team of AI agents doing research uses how many more tokens (word pieces) than a plain chat?

  1. About 2 times
  2. About 15 times
  3. About 150 times
Show the answer

About 15 times

Anthropic measured single agents at about 4x a chat and multi-agent systems at about 15x. On the BrowseComp test, token usage alone explained 80 percent of the variance in performance.

Source: Anthropic, 2025