Context is king as agents evolve, but problems persist

Byhassanhosting531@gmail.com

Context makes or breaks an AI agent.

It’s what gives AI agents the situational awareness to deliver trustworthy, accurate outputs. Without it, all other parts might be in place, but the relevant meaning that powers an agent to perform properly will be absent, dooming it to fail.

When provided with proper awareness, agents can help make medical diagnoses, personalize retail experiences, discover new clients, generate sales campaigns, manage and optimize entire supply chains and monitor vast swaths of data that are impossible for humans to oversee to detect fraud. Sometimes, they’re even building other agents.

But to deliver those outputs, agents need proper context. They need relevant, high-quality data and task-appropriate business logic, and the ability to discover and ingest them independently.

“It’s no different than providing context to an employee doing a job,” David Menninger, an analyst at ISG Software Research, told TechTarget. “You wouldn’t give a junior analyst a task without giving them some context.”

ISG Software Research analyst David MenningerDavid Menninger

Now, after the laborious task of getting data ready for AI and investing heavily in experimental pilots — while model reasoning capabilities improved — some enterprises are ready to put agents into production at scale and reap the benefits.

The most advanced organizations already have. But widespread deployment of multi-agent networks requires difficult work that could take years to complete, according to Menninger.

“We’ll … gradually eat away at the parts that are difficult, and we’ll hopefully, in a several year time period, have tackled the biggest challenges,” he said.

AI success amid failures

Despite the time and money that enterprises have invested in agentic AI development, most AI projects don’t make it into production.

Studies show that the failure rate is going down, which is evidence that enterprises that have addressed the underlying data issues that often stall AI projects are achieving at least some success connecting AI applications with the context they need to deliver trustworthy outputs.

But they also reveal that more AI projects still fail than succeed.

It’s no different than providing context to an employee doing a job. You wouldn’t give a junior analyst a task without giving them some context.
David MenningerAnalyst, ISG Software Research

For example, Deloitte’s 2026 State of AI in the Enterprise report, released in January, showed only one-quarter of the organizations surveyed have been able to move 40% of their AI experiments — whether single agents, multi-agent systems, chatbots, traditional machine learning projects or other pilots — into production. Similarly, Menninger noted that ISG’s year-end 2025 research found that about one-third of the 1,200 projects it examined are moving into production, an improvement from 17% a year earlier but still well under 50%.

“[Some] enterprises are no longer experimenting with agents,” Michael Bendersky, director of research at Databricks, told TechTarget. “They’re deploying them for real, complex work.”

Databricks’ own 2026 State of Agents report, published in January, found a 327% increase in usage of multi-agent systems built on domain intelligence over the previous four months, demonstrating that some enterprises — after there were almost none until late 2025 — have built at least remedial agentic systems.

Nevertheless, long-term challenges to building broad networks of context-aware agents remain.

Progress despite problems

One challenge to putting multi-agent systems into production is the completeness and quality of the AI workflow, according to Bendersky.

Humans used to perform all aspects of data preparation, retrieval and operationalization for analytics and AI, and they had time to ensure the quality of all parts of the pipelines that connect data with applications. Once they’re in production, agents are the ones calling on context, and for them to do so properly, everything they require for accuracy must be ready the instant they need it.

Enterprises that haven’t already done so need to modernize — and in some cases, overhaul — their data and AI infrastructures, which can be time-consuming and costly depending on how much needs to be revamped.

To connect agents with context, infrastructures need tools that improve data discovery to ensure agents call on context-appropriate information. Among others, they include the following:

  • Vector embedding and reranking models.
  • Semantic layers and data retrieval engines that are purpose-built for AI workloads rather than BI tasks.
  • Security capabilities that address new risks posed by agents.
  • Governance frameworks that oversee not only data but also agent behavior.
Michael Bendersky, director of research at DatabricksMichael Bendersky

“What will still remain difficult is the environment,” Bendersky said. “We are moving from an enterprise where most data access is done by humans to one where interactions with the data are done by agents. This will require continued investment in architectures where the right context is accessible to the agent at the right time. Governance, platforms, auditability, etc., will continue being challenging.”

In particular, semantic modeling capabilities are proving to be a crucial enabler for enterprises successfully building agents and need to be part of an organization’s AI workflow, according to Cindi Howson, chief data and AI strategist at ThoughtSpot.

Model reasoning capabilities have improved to the point that when large language models (LLMs) are asked questions that require them to call on limited amounts of contained data, they are 95% accurate. However, a September 2025 study demonstrated that when LLMs are asked to derive outcomes based on broad swaths of data spread across myriad systems, their accuracy falls to around 50% or lower.

A separate April 2026 study by DBT Labs — which provides semantic modeling capabilities — showed that accuracy again approaches 100% when LLM queries, often using traditional retrieval-augmented generation (RAG) pipelines to connect models with AI tools, are augmented by a semantic layer.

“Some organizations thought they could just use a basic RAG approach, and it’s not happening — it’s just not good enough,” Howson told TechTarget. “[They need] semantic layers and context layers, and humans in the loop to coach agents.”

The next challenge(s)

While known problems slowly get solved, new ones constantly emerge.

For example, connecting agents with context only became the dominant trend in data management and AI at the start of 2026 after unreliable model reasoning was viewed as the initial culprit holding back agentic AI development. Only once model reasoning capabilities reached a certain threshold and agents still weren’t consistently delivering accurate outputs did context become the primary focus, with a litany of vendors introducing tools to connect agents and context.

Now, data and AI infrastructures that don’t effectively connect agents with context are a principal focus, but as Menninger noted, it will take time to solve on a widespread basis. In the interim, Incremental improvements will be made that enable more enterprises than those now at the forefront to move agents into production and perhaps even build multi-agent networks.

“We’ll be making progress, identifying the bigger challenges and devoting research to solving those,” Menninger said.

But what then?

Assuming a combination of enterprise investment and technological advancement continues to result in a rising AI development success rate, managing systems that include thousands of autonomous agents will be a significant challenge, according to Donald Farmer, founder and principal of TreeHive Strategy.

Donald Farmer, founder and principal of TreeHive StrategyDonald Farmer

“Multi-agent systems — or swarms – [are one emerging issue],” he told TechTarget.

Agents within connected systems are meant to work together, to collaborate on a scale beyond human capacity. But for them to do so without overstepping boundaries, without exposing sensitive information, without accessing data they’re not authorized to operationalize, without violating regulatory statutes, without breaching data sovereignty rules — without behaving exactly as intended — is challenging. 

“Once individual agents have great context, [organizations have to figure out] how do you get an accounting agent, a legal agent and a procurement agent to collaborate, negotiate and hand off tasks to one another without human intervention or emergent anomalies,” Farmer said.

Another issue enterprises will have to address is how much responsibility to give agents and how much to keep under human supervision, according to Menninger.

He noted that ISG research shows that of those agents currently in production, 50% assist humans, 40% work in conjunction with humans but will eventually be empowered to act autonomously, and 10% are fully autonomous. Figuring out the right balance will be critical for enterprises to derive the benefits of agentic AI without suffering accidental consequences.

“There’s always going to be a next problem or opportunity,” Menninger said.

Eric Avidon is a senior news writer for Informa TechTarget and a journalist with more than three decades of experience. He covers analytics and data management.

Similar Posts

Leave a Reply

Your email address will not be published. Required fields are marked *