Difference between revisions of "AI Agent"

Revision as of 17:17, 26 August 2025

Definition

Context

Time is just a metric for elapsed opportunity reframes chronology as a ledger of potential, not inevitability. It strips away the metaphysical weight and instead treats time as a relational accounting system: not what is, but what could have been.

Here are a few angles:

• Time as Opportunity Delta
  • Time doesn’t measure existence—it measures missed or seized chances.
  • Every tick is a counterfactual ledger entry: what was possible, what was chosen, what was deferred.
  • In this framing, time is less about duration and more about decision density.
• Agentic Implications
  • For agentic systems, this view aligns beautifully: time becomes a negotiable affordance, not a fixed constraint.
  • Agents might track opportunity exposure rather than wall-clock time—e.g., “how many viable paths were available in this epoch?”
  • It invites temporal pluralism: different agents might experience time differently based on their opportunity bandwidth.

Access Authorization

Legacy Technology

Multiple Agent Access

Why AI agent security needs two distinct layers webinars at cerbos

Off the back of my MCP Webinar last week, I’ve been having a lot of conversations about authorization for AI agents - or really any workflow where multiple workloads are performing actions, directed by an LLM or otherwise.

There’s a growing consensus forming on what’s possible today versus what still needs deeper design work:

Layer 1: Basic service-to-service authorization (solved-ish)

Covers the foundational identity and access layer:

• User → MCP server
• Agent ↔ agent communication
• Agent → tool access

This is largely solved today. You can lean on existing frameworks like SPIFFE, OIDC, and service accounts to handle workload identity issuance and enforce authorization based on those identities.

Layer 2: Runtime transaction scoping (the hard problem)

This is where things get interesting and unsolved. It’s about scoping permissions down to a specific task or tool call within a single agent interaction.

• Token exchange is an early step here, but it doesn’t yet capture the concept of “this token only authorizes this task, for this transaction, and nothing else.” (Maybe RAR?)
• I’m currently experimenting with Transaction Tokens (credit to Atul Tulshibagwale) as a promising approach in this space.
• Macaroons are another option worth watching.

From what I’m seeing, most organizations are focusing on Layer 1 for now (correct, get your house in order) and leaving Layer 2 until IAM standards catch up (no clear architectural standard yet).

Future

• 2025-08 What's the Future of AI Language Models as a Decentralized Technology? ACLU

Sessions

Naming

Agent Name Service

Reference