Agentic Workflows

Agentic workflows are software systems in which a large language model decomposes a goal into sub-tasks, selects tools, executes actions against external systems, and revises its plan based on observed results — without per-step human prompting.

Why It Matters

Agents are the first credible path from "AI as autocomplete" to "AI as labor." Every SaaS category whose value is sequential clicks (CRMs, ITSM, RPA, ETL) is now contested by agent layers that operate them on a user's behalf.

Timeline

1. 2023-03

   AutoGPT and BabyAGI ignite the autonomous-agent narrative.

2. 2023-11

   OpenAI ships Assistants API with built-in tool use.

3. 2024-06

   Anthropic releases tool-use docs; agent stacks consolidate around JSON-tool calling.

4. 2024-10

   Anthropic Computer Use moves agents from APIs to actual screens.

5. 2025-01

   OpenAI Operator + Deep Research launch; "agent" becomes a product surface, not a library.

6. 2025-Q3

   Enterprise procurement language begins distinguishing "agent seats" from "user seats."

• 6 monthsmedium confidence

  At least one Fortune 500 reports an "agent FTE-equivalent" KPI in earnings.

• 12 monthsmedium confidence

  Browser-native agent runtimes ship in 3 of the top 5 browsers.

• 24 monthshigh confidence

  A vertical agent (legal, support, or sales) crosses $100M ARR purely on per-task pricing.

Related Trends

• Tool-Use Protocols

  Emerging

  Standardized ways for models to discover and call external tools (MCP, A2A).

• Computer Use

  Emerging

  Models that drive a real screen, mouse, and keyboard like a human.

• Reasoning Models

  Rising

  Models trained to think in long chains of intermediate tokens before answering.

• Inference Cost Collapse

  Mainstream

  Token prices falling 80–95% per year while quality keeps rising.