Agent projects fail around glue
A practitioner tracked agent projects and found most failures trace back to surrounding systems—unclear workflows, brittle context, poor tool contracts and weak operational design—rather than the agents themselves. The write‑up recommends treating integration, contracts and operational design as first‑class engineering work when deploying agents. (dev.to)
An artificial intelligence agent is a language model wired to tools and memory so it can take multi-step actions, not just answer one prompt. OpenAI says those systems need orchestration, tool execution, approvals and state outside the model itself. (developers.openai.com) That outside machinery is where many teams are now focusing their postmortems. In a recent practitioner write-up, Varun Pratap Bhardwaj argued that most failed agent projects break around workflow design, context handling, tool contracts and operations, not because the core model is incapable. (dev.to) The same pattern shows up across vendor documentation and field reports. Anthropic said on December 19, 2024 that teams succeeding with large language model agents usually rely on simple, composable patterns, while LangChain says the “right” context is more often the blocker than raw model capability. (anthropic.com) (docs.langchain.com) Context is the packet of instructions, history, tool outputs and external data the model sees on each turn. Anthropic wrote on September 29, 2025 that context is finite and must be curated across turns because an agent loop keeps generating new information that may or may not belong in the next call. (anthropic.com) Tool contracts are the other quiet failure point. OpenAI’s agents documentation says developers need to define specialist ownership, runtime behavior, results, state, guardrails and human review, which means the agent only works reliably when every tool has clear inputs, outputs and stop conditions. (developers.openai.com) That pushes agent work closer to classic systems engineering than chatbot prompting. LangChain describes the agent loop as repeated model calls plus tool execution, with middleware controlling what is passed in, what is saved, and what happens between steps such as summarization, logging and guardrails. (docs.langchain.com) Anthropic draws a line between workflows and agents that helps explain why some deployments stall. Its guidance says predefined workflows are better for well-defined tasks, while freer-form agents make sense when flexibility is needed, and many applications still work with a single model call plus retrieval instead of a full agent loop. (anthropic.com) The operational burden is now large enough that vendors are productizing it. Anthropic launched Claude Managed Agents on April 8, 2026 with sandboxing, orchestration and governance features aimed at shortening enterprise deployment work from months to weeks. (siliconangle.com) OpenAI’s current SDK docs make the same division of labor explicit: use the software development kit path when your server owns orchestration, state, approvals, integrations and observability. In practice, that means the “glue” is no longer side work around the agent; it is the product surface teams have to build first. (developers.openai.com)
