Agentic AI in the Enterprise 2026
In CXOTalk #916, Insight Partners MD Praveen Akkiraju describes an enterprise agent landscape moving at a "dizzying" pace — sophisticated enterprises have already deployed over a thousand agents in production, yet the governance, cost controls, and guardrails to run them safely are still being figured out in real time. Akkiraju frames the agent itself as "the harness" — the tools, context, memory, guardrails, observability, and evals around a stateless model — and warns that in multi-agent architectures errors compound as one agent feeds mistakes into the next. Host Michael Krigsman presses on agents that silently ignore instructions and confidently return wrong results.
The episode describes over a thousand agents already in production with data reached via untracked API/MCP/CLI paths and budgets blown in 90 days, while governance, guardrails, and cost controls are admittedly still being answered in real time. Human-in-the-loop is the only mature safeguard — a manual backstop, not a governance system.
Why governance matters
- Agent sprawl is already here: Akkiraju says "sophisticated enterprises today have deployed over a thousand agents in production" — and openly asks how you ensure they "have access to the right tools" and are "governed properly," the exact control gap governance exists to close.
- Untracked data access via API / MCP / CLI: an agent accesses data by "either giving it an API or… access to an MCP server, or maybe you're using CLI" — each with different granularity, so nobody can attest to what agents can reach without deliberate interface governance.
- Multi-agent error compounding: "an agent makes an error and feeds it into another agent, you essentially have a compounding effect of errors" — governance requires granular, per-step observability of traces, not just checking the final output.
- Token/cost governance is now a hard control: a Goldman Sachs study cited in-episode found enterprises burning "a year's worth of AI budget in 3 months" via "token maxing," forcing CIOs to prioritize projects and weigh cost against ROI.
- Guardrails and evals are the safety layer, and they're hard: Akkiraju calls the harness — "clear guardrails, clear constraints, and clear evaluation criteria" plus sandboxes and observability — a critical concern, while conceding the evals are "one of the more challenging aspects."
- Silent, confident failure demands human oversight: Krigsman warns an agent can ignore instructions and admit a wrong answer seamlessly — "there's no way to know unless you know something about the answer" — which is why human-in-the-loop remains the paradigm.
- Third-party agents are a supply-chain risk: on public agent libraries Akkiraju cautions "be very careful about deploying a third-party agent into your environment unless you truly have confidence… we're still not at that level of maturity."
In their words
This is notion of token maxing which is as many tokens as you provide will get consumed as quickly as possible.— Praveen Akkiraju
You have a multi-agent architecture and an agent makes an error and feeds it into another agent, you essentially have a compounding effect of errors.— Praveen Akkiraju
When you look at the way an agent accesses data, you're either giving it an API or you're giving it access to an MCP server, or maybe you're using CLI.— Praveen Akkiraju
There's no way to know unless you know something about the answer that you've been given. This is very dangerous.— Michael Krigsman