Looking Back at 2025: The Year We Found the Edges

What we learned about AI agents from the trenches - a practitioner's perspective

Remember "The Year of the Agent"?

It's New Year's Eve, and I've been thinking about all those headlines from this time last year claiming 2025 would be the "Year of the Agent."

At the end of 2024 and into early 2025, you couldn't move in tech media without hitting a prediction that "2025 will be the Year of the Agent." VC theses, conference talks, blog posts - all saying the same thing: autonomous AI agents were going to transform knowledge work, orchestrate complex workflows, and quietly replace a big chunk of coordination and glue work in our organizations.

A year later, if you're running infrastructure for a mid-market company - especially outside Silicon Valley - it probably doesn't feel like that happened. You might have Copilot, a few pilot projects running, and some impressive demos. But very little you'd trust to run unattended at 3am.

Here's what I think actually happened in 2025:

2025 wasn't the Year of the Agent. It was the year we discovered what agents can't do reliably yet - and quietly built the infrastructure that might make them work later.

The irony isn't lost on me: this article itself was assembled with the help of multiple AI agents - search, summarization, comparison - stitched together by hand. The process worked, but it was brittle and deeply personal. I tweaked prompts, rewired tools, and patched failures on the fly.

At the end of 2025, I think people have been able to get agents to do amazing things, but it's still all very custom. Still all very "Excel wizard" territory.

Where Agents Actually Are (Production Reality)

When you look past vendor decks and into production surveys and incident reports, a very different picture emerges:

The Numbers Don't Lie

80% of companies say their AI agents have taken unintended actions - including accessing unauthorized systems, exposing sensitive data, or downloading sensitive content inappropriately. (SailPoint 2025 survey of 100+ security leaders)
95% of AI agent projects failed to reach meaningful production in 2025. MIT surveyed 1,837 teams; only 95 were actually running agents in production. (MIT 2025 study)
70% of teams with agents in production rebuild their AI stack every three months or faster - framework churn and breaking changes make stability nearly impossible. (Cleanlab AI Production Survey 2025)
96% of enterprises plan to expand agent use, but only around 1% report systems they're comfortable letting make independent decisions. (McKinsey 2025 survey)

That last point is the paradox of 2025: nearly everyone wants more agents, but almost nobody trusts them.

This is normal in technology adoption. Every disruptive wave has a 3-4 year "trough of disillusionment" where we discover all the edge cases and failure modes the hype never mentioned.

Cloud went through it in 2008-2011 - remember "the cloud is just someone else's computer" and all the security hand-wringing? Virtualization went through it in the early 2000s when VMware ESX was still figuring out memory overcommitment and live migration quirks.

AI agents are right in that trough now.

The Terminology Retreat

One of the clearest signals that the reality of agents isn't as rosy as the headlines led us to believe is the language shift over the last 18 months. We've quietly walked back the boldest words:

"Autonomous agents" became "Agentic workflows"
"Fully automated" became "Human-in-the-loop"
"Agent systems" became "AI copilots"
"Self-improving" became "Guided learning"

That's not just marketing spin - it's the industry subconsciously admitting what ops and infra people already know:

These systems are powerful, but they are not robust enough to be left alone.

We're building assistive automation, not digital employees. At least, not yet.

Enterprise Reality: Copilot, Stack Churn, and the Excel Wizard Problem

Copilot in the Real World

At the start of 2025, Microsoft added "agent functionality" inside Copilot across Microsoft 365. On paper, it's compelling: summarize meetings, draft emails, pull numbers from Excel, eventually orchestrate tasks.

In practice, what I hear - and experience - is more mixed:

Confusion about where data goes and how it's stored
Lots of "almost right" outputs that still require human verification
Licensing and rollout complexity that outpaces the value for many mid-market orgs
Scaffolding the tools and context for agents to generate quality output requires significant technical expertise

Copilot is agentic under the hood, but from an infra perspective it behaves more like a noisy junior assistant than a reliable operator.

The Stack Problem

One of my favorite stats from the research: 70% of teams running AI agents in production rebuild or significantly reconfigure their stack every three months or faster.

That matches my lived reality. My own "agent stack" is in constant flux:

New orchestration frameworks
New observability tools
New safety layers
New model choices as pricing and quality shift

Every "shiny new thing" promises to finally make agents robust and manageable. Every time, I have to choose between learning yet another framework or finishing work that actually pays bills.

"Managing Agents" as the New Excel Skill (and Why That's a Problem)

Here's a useful analogy:

Every company has one or two "Excel wizards" - people who could bend spreadsheets into full applications. Those skills were unevenly distributed. Most staff could do a VLOOKUP, but only a few understood the monster workbook that actually ran payroll reporting.

We're in the same place with agents:

A small number of people can make agents do magic - multi-step workflows, tool orchestration, solid prompting, debugging
Most people try a few times, get inconsistent results, and retreat to familiar tools

The skill gap matters because current-gen agents are not "set and forget." They require:

Continuous prompt and tool refinement
Careful scoping and permissioning
Monitoring, evaluation, and feedback loops

And that's before cost enters the picture.

On my own Claude $30/month plan, I routinely slam into rate limits as soon as I start running longer agentic workflows (research, planning, coding, and retrieval chained together). In fact, I had to add credits to my Anthropic account just to finish the research for this post. To keep things smooth, I need to move into higher-tier or Opus-based usage, which can get expensive fast if you're running agents all day across a team.

For most individuals and SMBs, that pushes "serious agent use" back into specialist hands - the AI equivalent of the Excel wizard - rather than democratizing it.

Where Agents Do Shine: Infrastructure and Config Work

One place I've found agents genuinely transformative is infrastructure and configuration management - not application development or user-facing features.

Why? Because infrastructure work has two great properties for LLMs:

Strict syntax and structure. Config files, router ACLs, firewall rules, Docker Compose, YAML - all of it has to be exactly right.
Clear feedback loops. Either the container comes up, the port opens, the service responds... or it doesn't.

Anyone who's fought with YAML and JSON knows that a single misplaced space or brace can break an entire pipeline. This is precisely the kind of pattern-heavy, highly structured task LLMs are surprisingly good at.

The FabLab Example

In my own "FabLab" at home - a personal datacenter I'm building out with leftover servers running open-source infrastructure software like Proxmox - I've been testing this. With my agent by my side, I was able to:

Describe the financial tooling I wanted (Firefly III with a Postgres backend, proper backups, and reverse proxy)
Ask an infra-focused agent to:
- Design a Docker-based architecture
- Generate the Docker Compose and .env files
- Document the whole setup in a project folder using git to monitor changes

With a bit of review and small tweaks, I had the system up and ready for real data inside an hour. And with every deployment it gets better because I can refine the templates - Bob, my personal agent working alongside me makes that possible.

The time savings weren't in "creative insight" - they were in removing 200 lines of brittle, repetitive YAML work and documenting it as I went.

What This Means for Infrastructure Work

I'm not suggesting we let agents make changes to production systems unsupervised. But for tasks like:

Security reviews on large config files
Quick documentation updates
Generating infrastructure-as-code templates

...agents make those workflows dramatically more accessible.

That said, certain automations might be workable if we're careful and move slowly. In my lab, I've built a Claude Code skill that lets Bob add DNS records on the fly using the OPNsense firewall API - but only when I explicitly approve each change.

The promise of infrastructure-as-code has always been hampered by the learning curve and syntax complexity. Having an agent that can hold the context of your entire environment while writing configs and templates is genuinely helpful.

That's the key pattern I see for infra and ops:

Treat agents as syntax accelerators and documentation engines for your environment, not as fully trusted, on-call reliability engineers you hand the pager to.

Used this way, they help chip away at the IT debt that accumulates as systems sprawl - without pretending they can own your incident queue.

Security & Privacy: Bob and Local Models

I spend a lot of time thinking about security and AI. Bob, my own personal agent and "business partner" is built on top of Claude Code and Daniel Miessler's PAI repository - an open-source framework that lets you build context-rich AI assistants with persistent memory, custom skills, and workflow automation. Bob helps with:

Financial tracking and forecasting
Drafting proposals and quotes
Managing marketing content and basic ops tasks

I wrote more about how Bob has helped me in "Meet Bob: 48 hours with my AI business partner" on my site, but the important point is this:

Bob is genuinely useful because he has deep context about my life and business. That's also what makes him dangerous if mishandled.

Am I comfortable trusting Anthropic with my personal and business data? For now, yes - I'm trusting their SOC2 compliance status and the fact that they're HIPAA-compliant. If they can handle protected health information, I figure they can handle my financial forecasts and client proposals.

The Broader Security Picture

The data bears out the risk side:

69% of organizations say they've deployed AI agents, but only 21% have adequate visibility and controls to secure them - a 79% blind spot. (Akto Security Survey 2025)
39% of companies reported agents accessing systems they weren't supposed to, and about a third saw agents touch sensitive data inappropriately. (Akto Security Survey 2025)

For infra and security teams, nothing about this is mystical. The controls we need are the same ones we've always needed:

Least privilege for tools and agents. If an agent doesn't need shell access, don't give it shell access.
Segmentation and blast-radius control. Run agent-connected services in constrained environments, with clear egress controls.
Auditability. Log every tool call and decision as if a junior admin were doing it.
Human approval for high-risk actions. Treat "human-in-the-loop" not as a marketing phrase, but as a control requirement.

Running Models Locally

For highly sensitive data - PII, PHI, law enforcement data, deeply personal context - it's time to take local and open-weights models seriously.

Tools like Ollama and LM Studio make it straightforward to run powerful AI models on your own hardware. We can run models like the just-released DeepSeek 3.2, Mistral, and Qwen locally, accepting the operational burden (GPU requirements, model management, slower inference) in exchange for complete data control.

For infrastructure work where you're dealing with network diagrams, security configs, or sensitive architecture documentation, keeping that data on-premises isn't paranoia - it's good practice.

So... Was 2025 the Year of the Agent or Not?

There Was Real Progress (Q4 2025)

To be fair, there was real infrastructure progress this year, especially in Q4:

Anthropic shipped the Claude Agent SDK (September 29, 2025), moving from "cobble your own" to a more structured agent framework
Claude Code hit $1B+ in revenue (December 3, 2025), proving there's a real business around agent-assisted development
Snowflake announced a $200M partnership (December 3, 2025) focused explicitly on agentic AI
Accenture committed to multi-year agent deployments (December 9, 2025) for Fortune 500 clients
Microsoft expanded deep integration of Claude into Copilot and GitHub (November 18, 2025)

If your definition of "Year of the Agent" is "the year the infrastructure finally got serious", then 2025 arguably qualifies.

But That's Not What the Hype Promised

Those early-2025 headlines were about:

Autonomous agents transforming day-to-day work
Fully automated workflows replacing coordination roles
Agents reliably handling complex, multi-step tasks without supervision

On that definition, the research is clear:

Agents remain experimental, heavily scaffolded systems that require close supervision. Most organizations are still stuck in pilot purgatory. The people getting value are those treating agents as constrained assistants with strong guardrails, not as autonomous coworkers.

As Andrej Karpathy has pointed out, LLMs are astonishing but probabilistic - they're sometimes wrong and genuinely hard to compose into reliable systems. Building production systems around them is as hard as building early distributed systems. This is going to take years, not quarters.

Curtis Northcutt (CEO of Cleanlab) has said publicly that agentic AI probably won't really "gel" until around 2027, once we have better tooling, patterns, and operational experience.

That aligns almost perfectly with the historical 6-10 year maturation cycle we saw with cloud and virtualization.

If You're Frustrated, You're Probably Doing It Right

One of my favorite insights from the research is:

"If you're frustrated with AI agents right now, that means you're doing the real work."

The "real work" here is:

Instrumenting agents like any other distributed system
Designing for failure modes, not happy paths
Narrowing scopes until workflows are testable and debuggable
Writing runbooks and roll-back plans instead of just demos
Documenting what doesn't work, not just what plays well on stage

The hype cycle's trough isn't a failure - it's where we separate technology that works from technology that merely demos well.

The practitioners who stay in the game now - infra engineers, sysadmins, SREs - are the ones who will define the patterns everyone else uses when this finally stabilizes.

Why We Need a GrayBeard AI Collective

Back in November, I met up for beers with two colleagues I've known for years. Both are veteran IT infrastructure people - the kind who've been through enough hype cycles to know the difference between demos and 3am pages.

We got talking about Bob, my AI business partner I built on Claude Code and Daniel Miessler's Personal AI Infrastructure (PAI). They were curious about the architecture, how I'd structured the context, what actually worked versus what was still brittle.

The conversation drifted to the old Manitoba VMUG - the VMware User Group that used to bring infrastructure people together. Sure, there were vendor presentations (that's how it was paid for), but they were technical deep-dives, and the real value was practitioners showing what actually worked in production and sharing honest war stories.

"We should have a place like that to share what works with AI," one of them said.

A few days later, I had lunch with a couple of colleagues. Same themes - agents, infrastructure reality, what actually works. That's when the joke started: if this lunch was the "first meeting," then the next one would be the second meeting of the GrayBeard AI Collective - where I'd show off Bob and what AI can actually do for infrastructure work.

GrayBeard doesn't mean you literally have gray hair (though in my case it's true). In infrastructure circles, it means you've been in the trenches long enough to see the deeper patterns in systems. You know what it looks like when hype meets operational reality. You've deployed enough "revolutionary" technologies to recognize when something's ready versus when it's still baking.

This is exactly why that conversation stuck with me.

We need a place for:

Infra and ops people who've lived through previous hype cycles
Folks actually wiring agents into real systems, not just building demos
People willing to share failures, ugly hacks, and half-working patterns so the rest of us don't repeat them

If you're feeling overwhelmed by the pace of AI development, or worried you'll be left behind if you don't keep up - you're not alone. That conversation over beers confirmed it. Lots of experienced people are in that same place.

Maybe you've played with Copilot at work, or spun up your own AI experiments at home. Maybe you're the person everyone asks "what's the deal with this AI stuff?" because you're the one who actually understands infrastructure.

That's who this is for.

Not vendor pitches. Not slideware. Not "thought leaders" selling courses.

Just practitioners sharing what actually works - and being honest about what doesn't.

Join the Second Meeting

The first meeting was that conversation in November over lunch where we talked about AI agents and infrastructure reality.

The second meeting is about showing, not just talking. I'll demonstrate Bob's infrastructure capabilities, share what's actually working in my FabLab, and hear what you're building.

If you're:

The person everyone asks about Copilot
The one maintaining the homelab that quietly outperforms half your vendor stack
The sysadmin trying to balance "play with the new thing" against "please don't break payroll"

...then you're exactly who this is for.

Connect with me:

Comment below
Find me at wallykroeker.com
Email: [email protected]

2025 wasn't the Year of the Agent.

But if we do the boring, disciplined work now - documenting failures, sharing patterns, building guardrails - we might look back in 2028 and realize this was the year the right people started taking agents seriously.

Sources

SailPoint 2025 Survey - 80% unintended actions (Technology Magazine)
MIT 2025 Study - "The GenAI Divide: State of AI in Business 2025" - 95% project failure rate (1,837 respondents, 95 in production) (Fortune coverage)
Cleanlab AI Production Survey 2025 - 70% rebuild stack every 3 months (cleanlab.ai)
Akto Security Survey 2025 - 69% deployed, 21% visibility, 39% unauthorized access (akto.io)
Anthropic Official News - Claude Agent SDK (Sept 29), $1B revenue (Dec 3), Snowflake $200M (Dec 3), Accenture partnership (Dec 9), Microsoft integration (Nov 18)
Daniel Miessler's Personal AI Infrastructure - GitHub Repository

Full bibliography and research methodology available on request.

About the Author

Wally Kroeker is a security and infrastructure practitioner with 20+ years experience. He runs GoodFields Consulting, focusing on security architecture and pragmatic AI implementation. He writes about technology, infrastructure, and building humane systems at wallykroeker.com.