If you've been following the AI agent narrative, you've
probably heard the promises: autonomous systems that can handle your trading,
book your flights, monitor competitors, and manage entire workflows without you
lifting a finger.
And yet.
A Gartner projection from June 2025 suggests that over
40% of agentic AI projects will be canceled before the end of 2027, citing
escalating costs, unclear business value, and inadequate risk controls.
Why?
Because the infrastructure hasn't caught up. The web was
built for humans, not for bots. And the gap between what agents can do in a
demo and what they can do in production is still enormous.
The Four Failures of AI Answers
Yaniv Tal, founder of Geo and co-founder of The Graph, has
spent a lot of time thinking about this problem. His diagnosis? The
unreliability of AI responses isn't just a model problem—it's a data problem.
He points to four core weaknesses in how information
circulates online:
1. Lost provenance. Content gets scraped,
republished, and re-scraped until the original source disappears. Models learn
a claim but have no clear path back to who first made it.
2. Flattened authority. Research papers, press
releases, and anonymous forum posts all end up in the same training pipeline,
treated as equivalent evidence.
3. Hidden disagreement. Language models compress
competing views into a single confident-sounding answer, erasing real disputes
among experts.
4. Repeated model-generated errors. When AI
outputs become training data for later models, mistakes get amplified across
generations.
"AI doesn't have a truth problem, the internet
does," Tal said.
A 2024 Nature study documented this exact
dynamic in action. Successive model generations trained on prior model output
progressively lose fidelity to the original human-produced data distribution—a
phenomenon researchers call "model collapse."
Why the Web Hates Bots
Even if the underlying models were flawless, the web itself
creates enormous friction for autonomous agents.
Consider what a capable agent actually needs: consistent
access, real-time responsiveness, and structured data it can parse without
human intervention. The web, as currently built, systematically discourages
automated access.
Oxylabs' upcoming Web Openness Index illustrates the gap:
- The
global average score for practical reachability—how well a
site responds to standard automated HTTP requests—stands at 83.4
out of 100.
- The
score for anti-automation friction—CAPTCHAs, rate limiting,
fingerprinting, bot detection—is only 62.8.
- Structured
data interoperability—whether sites return data in formats machines
can actually work with—drops to 60.3.
Those 20-plus-point differences reflect a structural gap.
Sites generally respond to requests for automated access, but restrictions
abound, and data is often returned in machine-unfriendly ways. Agents that
depend on reliable, timely, structured information will consistently fall into
that gap.
Inside organizations, agents face a related problem: data
that hasn't been cleaned, tagged, or structured in a way that AI can
understand. The relevant information exists—it's just not accessible.
The Data Access Crisis
Research from Cloudera and Google/MIT points to the same
bottleneck: AI agents can't reliably act on data they can't access,
understand, or retrieve in real time.
Cloudera's survey of 1,500 enterprise architects and cloud
infrastructure leads found that 95% had delayed or canceled AI projects in
the past year due to issues with data governance, compliance, or regulatory
problems. A wide majority said their current data architecture requires a
"significant overhaul" to meet AI goals.
"Even if enterprises are ready to use AI, many are
coming to the realization that the foundational infrastructure it relies on is
not," the report noted.
Similarly, more than half of 300 IT execs responding to the
Google/MIT survey said they have paused or delayed AI agent deployments to
address foundational data issues—entrenched silos, inaccessible data,
insufficient real-time access, and lack of context.
These are not minor complaints. They're the difference
between "we can try this" and "we can scale this."
The Security Problem
Maybe the biggest bottleneck is security. A Docker survey of
800+ developers and decision-makers found that 40% of respondents cite
security as their top blocker when building agents.
The problem isn't confined to a single layer of the stack:
- Infrastructure: As
organizations scale agent deployments, teams emphasize the need for secure
sandboxing and runtime isolation—even for internal agents.
- Operations: More
tools, more integrations, and more orchestration logic create blind spots.
Over a third of respondents report challenges coordinating multiple tools,
and a comparable share say integrations introduce security or compliance
risk.
- Governance: 45%
of organizations say the biggest challenge is ensuring tools are secure,
trusted, and enterprise-ready.
The Model Context Protocol (MCP) is widely adopted—85% of
teams further along in their agent journey are familiar with it, and two-thirds
actively use it across projects. But most teams are operating in what could be
described as "leap-of-faith mode," adopting the protocol without the
security guarantees they would demand from mature enterprise infrastructure.
For teams earlier in their agentic journey, 46%
identify security and compliance as the top challenge with MCP.
Organizations are increasingly watching for threats like prompt injection, tool
poisoning, and the more foundational issues of access control and
authentication.
"AI agent security is what sets the speed limit for
agentic AI in the enterprise," the Docker report concluded.
The Execution Gap
In crypto, the bottleneck is even more visible. An agent can
analyze market trends and generate strategies, but if it can't place orders,
manage positions, or interact on-chain, its analysis remains theoretical.
Gate.io's analysis
identifies three major execution obstacles for AI agents in the crypto
ecosystem:
- Interface
fragmentation. The crypto ecosystem spans centralized exchanges,
DEXs, wallets, and on-chain data, each with its own API standards and
authentication methods. Integrating and maintaining each connection is
time-consuming and expensive.
- Permission
and security risks. Autonomy requires access to trading systems
and assets. But prompt injection attacks, malicious plugins, API key
abuse, or automated errors can turn a system failure into real financial
losses. Industry reports reveal that 72% of enterprises are
operating AI agents without adequate risk controls.
- Lack
of standardized protocols. Most agents rely on different agent
frameworks and orchestration systems, but communication between modules
isn't standardized. Each integration requires custom
adaptation—fundamentally limiting scalability.
This is where crypto and AI intersect most directly: the
execution layer is the missing piece. Without it, agents remain analysis tools
rather than active entities.
What Needs to Change
Tal's proposed solution—through his startup Geo—is to
rebuild how information is structured at the source. His system separates
claims, authors, and evidence, preserving relationships between them.
Communities called "Spaces" would curate knowledge, ranking arguments
by assessed strength while keeping competing perspectives visible beneath the
top results.
"Pluralism doesn't have to mean noise," Tal said.
"We think several well-structured competing perspectives is often the more
honest answer."
The approach doesn't eliminate the hard questions—how
communities select editors, how they resolve manipulation, or whether expertise
in one subject should carry weight in another—but it offers a direction.
On the infrastructure side, the answer is less sexy but more
urgent: data needs to be accessible, structured, and secure. The web needs to
work for machines as well as humans.
This article is for informational purposes only and does
not constitute investment advice.
0 Comments