Agent Pulse
Posts
2025 in Agents: 10 Things That Actually Mattered

2025 in Agents: 10 Things That Actually Mattered

O P
December 31, 2025

In partnership with

AIonFire | AArena | Submit Agent | Advertise

Welcome back!

Most year-end takes will tell you 2025 was “the year of agents.”

That’s lazy.

2025 was the year agents stopped being a concept and started being constrained systems that either worked — or didn’t.
No fireworks. No AGI moments. Just a lot of quiet, important shifts.

Here are the 10 things that actually shaped the agent ecosystem this year.

Stop Planning. Start Building.

End of the year? Or time to start something new.

With beehiiv, this quiet stretch of time can become your biggest advantage. Their platform gives you all the tools you need to make real progress, real fast.

In just days (or even minutes) you can:

Build a fully-functioning website with the AI website builder
Launch a professional-looking newsletter
Earn money on autopilot with the beehiiv ad network
Host all of your content on one easy-to-use platform

If you’re looking to have a breakthrough year, beehiiv is the place to start. And to help motivate you even more, we’re giving you 30% off for three months with code BIG30.

Start building for 30% off today

1) Agents Entered Production as Background Systems

Waymo deployed Google’s Gemini as an in-vehicle agent handling passenger interaction, cabin controls, and trip context — explicitly separated from driving logic. This wasn’t a demo; it shipped inside a live robotaxi fleet.

Why this matters:
Agents entered production without branding. They became invisible infrastructure.

Example: Waymo testing Gemini as an in-car agent

2) Narrow, Task-Bound Agents Outperformed General Ones

Klarna publicly reported that its agents focused on specific commerce and support workflows, not open-ended autonomy — replacing hundreds of customer service roles while staying tightly scoped.

Why this matters:
The highest-ROI agents in 2025 were intentionally limited.

Example: Klarna’s internal AI agents

3) Agent Evaluation Shifted to Task-Level Benchmarks

Z.ai released GLM-4.7 with explicit agent-style benchmarks (BrowseComp, τ²-Bench, Code Arena) focused on tool use and multi-step task completion, not chat quality.

Why this matters:
Agents started being measured like systems, not conversations.

Example: Z.ai GLM-4.7 benchmarks

4) Agent Marketplaces Became Enterprise Infrastructure

Oracle launched an enterprise marketplace for partner-built, certified AI agents embedded directly into Fusion Applications — with security, vetting, and support guarantees.

Why this matters:
Discovery moved from social proof to institutional distribution.

Example: Oracle Fusion AI Agent Marketplace

5) Tool Reliability Beat Raw Model Intelligence

Anthropic open-sourced Agent Skills, standardizing how agents interact with tools from partners like Atlassian and Figma — focusing on predictable execution, not smarter reasoning.

Why this matters:
Tool correctness became more valuable than clever prompts.

Example: Anthropic’s Agent Skills standard

6) “Agentic” Was Redefined as Supervised Autonomy

Microsoft emphasized human-in-the-loop checkpoints, auditability, and bounded autonomy in Copilot Studio, moving away from fully autonomous positioning.

Why this matters:
The industry accepted that unchecked autonomy is a liability.

Example: Microsoft Copilot Studio agents

7) Enterprises Adopted Faster Than Startups

SOMPO Holdings deployed internal AI agents to ~30,000 employees, standardizing workflows across insurance operations.

Why this matters:
Once risk was controlled, enterprises scaled faster than startups.

Example: SOMPO AI Agent rollout

8) Multi-Agent Systems Stayed Mostly Experimental

Despite heavy research output, most production deployments in 2025 avoided large agent swarms due to debugging, cost, and evaluation complexity.

Why this matters:
Coordination remains unsolved at scale.

Example: Academic & enterprise agent research

Source: Stanford WORKBank / MIT Sloan research

9) Cost Became a Core Design Constraint

As inference costs scaled, agent teams redesigned systems around bounded reasoning, selective tool calls, and fallback logic to keep per-task costs predictable.

Why this matters:
Agents became economically engineered systems.

Example: OpenAI & Anthropic pricing pressure

10) Agents Began Competing With Software, Not Chat

Across finance, support, and research, agents increasingly replaced scripts, dashboards, and manual ops, not chat interfaces.

Why this matters:
That’s when agents stopped being “AI features” and became economic actors.

Example: Internal enterprise agents replacing SaaS workflows

AI-native CRM

“When I first opened Attio, I instantly got the feeling this was the next generation of CRM.”
— Margaret Shen, Head of GTM at Modal

Attio is the AI-native CRM for modern teams. With automatic enrichment, call intelligence, AI agents, flexible workflows and more, Attio works for any business and only takes minutes to set up.

Join industry leaders like Granola, Taskrabbit, Flatfile and more.

👉Start for free today

✨ Featured Agents You Shouldn’t Miss