AI agents are entering a different phase.

They are no longer limited to generating text or answering questions. They can research, manage files, interact with websites, communicate across platforms and execute multi-step workflows. But as agents become capable of acting on our behalf, I think one question becomes increasingly important:

How do we give an AI enough autonomy to be useful without giving it unrestricted authority?

That is what makes IronClaw 1.0 from #NEARAI particularly interesting to me.

Separating Thinking From Acting

IronClaw approaches agent security at the architectural level.

Instead of allowing the model that makes a decision to directly execute the resulting action, #ironclaw places a guard layer between decision-making and execution. Actions pass through this controlled path, and sensitive operations can require explicit approval before they are carried out.

That creates a useful separation:

The agent can decide what it wants to do, but it doesn't automatically have unrestricted permission to do it.

For me, that's more interesting than simply adding another security feature to an AI agent. The control mechanism becomes part of how the agent operates.

#ironclaw also tackles another practical problem: what happens when an autonomous workflow gets interrupted?

Its continuous checkpointing allows work to resume instead of forcing the agent to start again. Persistent state and memory can also carry across CLI, Web, Slack and Telegram, allowing the same assistant to retain context across different interfaces.

For organizations, #ironclaw supports shared tools and skills through multi-tenant deployments while maintaining workspace boundaries, with single-tenant deployments available where complete isolation is required.

The Performance Behind the Architecture

The security model would be less compelling if it came at the expense of capability.

In NEAR AI's July 27 evaluation of Ironclaw 1.0, the system took the top reported position across three different agent benchmarks using the same DeepSeek-V4-Flash base model for the harness comparisons.

IronClaw recorded 93.5% on PinchBench, which evaluates 147 real-world tasks including scheduling, email triage, coding, research and file management.

On ClawBench, which tests multi-step tasks across more than 140 real production websites, IronClaw recorded 88.6%.

And on OfficeQA, which evaluates grounded reasoning over a large collection of U.S. Treasury documents, it achieved 76.4%.

These numbers should not be treated as directly comparable because each benchmark measures a different capability. What stands out to me is that the same architecture performed strongly across practical task execution, web interaction and document reasoning.

The Bigger NEAR AI Picture

IronClaw sits within the broader vision of #NEARAI , which focuses on private, confidential and verifiable AI infrastructure.

That brings the role of @NEAR Protocol and staking into the conversation.

Staking is often viewed primarily through the lens of rewards. At the protocol level, however, staked NEAR supports the validator infrastructure responsible for maintaining network consensus and security.

There is now an even more direct connection to AI infrastructure. #NEARAI has introduced staking-based access that allows users to stake NEAR for credits supporting confidential inference and always-on IronClaw agent hosting.

That changes how I look at the relationship between staking and decentralized AI.

The future isn't simply about building agents that can think better. It is about creating infrastructure where agents can act within boundaries, preserve their work, protect sensitive computation and operate on decentralized systems with meaningful security guarantees.

IronClaw 1.0 therefore represents something I find more valuable than another AI capability demo: a move toward making controlled autonomy a fundamental part of how AI agents are designed.

The question for the next generation of AI may not be how autonomous can agents become?

It may be:

How do we make autonomy trustworthy enough to use?