Oh, your agents need OAuth2, too

Summary of Oh, your agents need OAuth2, too

by The Stack Overflow Podcast

29m•November 3, 2026
Listen to Episode

Overview of Oh, your agents need OAuth2, too

This Stack Overflow Podcast episode, recorded at the LDX3 conference, explores how OAuth 2.0 and related identity standards are evolving for the “agentic era” of AI. Host Ryan Donovan talks with Dan Moore of FusionAuth about how AI agents should authenticate, how much access they should have, and why identity, security, and least-privilege controls matter even more when software can act autonomously.

Main Themes

OAuth 2.0 is still foundational

  • OAuth 2.0 remains a core building block for modern authorization, even as AI systems change how software interacts with identity and permissions.
  • Newer AI-related standards, especially Model Context Protocol (MCP), are increasingly being built on top of OAuth rather than replacing it.
  • The standards are evolving together:
    • OAuth is being adapted for agent use cases.
    • MCP is also pushing the ecosystem toward new identity-related standards.

Agents need their own identity

  • A major takeaway: AI agents should not simply “borrow” a human’s identity by default.
  • Dan argues that agents need:
    • Their own identity
    • Clear, deterministic permissions
    • Strong auditability
  • Letting agents run with a human’s credentials makes it difficult to track actions, limit damage, and understand accountability.

Human-delegated vs. autonomous agent access

  • There are two broad use cases:
    1. Humans delegating authority to an agent
    2. Agents acting more independently, like service accounts
  • For delegated access, OAuth’s redirect and login flow still matters.
  • For autonomous agents, the emphasis shifts more toward:
    • Bounding what the agent can do
    • Downscoping permissions
    • Sandboxing and environmental controls

Security and Access Control Best Practices

Default to least privilege

  • Dan stresses that organizations should move away from default-most-privilege models.
  • Better patterns include:
    • Time-boxed permissions
    • Single-use approvals
    • Downscoped tokens
    • Gateways or intermediary services that limit what an agent can access

Avoid “YOLO” agents

  • The hosts discuss how some users are letting agents operate too freely, including browser-driving agents and agents reusing human credentials.
  • Dan calls out the risk of treating AI as if it can simply be told “don’t break anything.”
  • Instead, teams should design systems where the agent’s actions are constrained in a deterministic way.

Sandboxing matters

  • Sandboxing is recommended, especially for agents with broad or risky capabilities.
  • Dan expects more agent execution to move to the cloud, where networks and runtime environments can be constrained more easily.
  • Running an agent on a local machine may be acceptable for experimentation, but cloud-based controls are likely safer and more scalable.

AI in Software Development: Survey Insights

Many teams lack measurement plans

  • In a survey conducted with LeadDev, 69% of respondents said they did not have a plan for measuring AI effectiveness in software development.
  • Dan notes that teams are often measuring what is easiest to measure, such as:
    • Token usage
    • Lines of code
    • Pull requests
  • But these metrics are imperfect proxies for actual value.

Developers care more about performance than cost

  • Another survey finding: most developers switch AI models primarily for performance, not cost.
  • Cost matters, but for high-leverage work like software development, better performance often justifies higher spend.

Signs of trouble in AI-assisted development

  • Survey data and anecdotal experience suggest:
    • Debugging ability may be declining on some teams
    • More developers are letting AI agents reuse human credentials
    • Large AI-generated PRs are creating “zombie review” problems

Cultural Shifts in Engineering

Engineering is still engineering

  • Dan argues that AI does not remove the need for real engineering.
  • The hard parts remain:
    • Understanding requirements
    • Shaping systems
    • Managing tradeoffs
    • Ensuring reliability and scalability
  • The hype cycle may change, but engineering fundamentals do not.

AI changes the role of the developer

  • Software developers may shift from hand-crafting every line of code to:
    • Defining the environment
    • Setting constraints
    • Reviewing outputs
    • Shaping system behavior
  • This may feel like a loss of craftsmanship for some, but Dan sees it as an opportunity for engineers who care about solving real business problems.

Smaller, reviewable work still matters

  • The conversation highlights frustration with massive AI-generated pull requests.
  • A key norm emerging: just because AI wrote it doesn’t mean it should be accepted as-is.
  • Teams should push back on poor structure and ask AI to break work into smaller, reviewable pieces.

Build vs. Buy: Identity and Boring Software

Don’t reinvent identity

  • Dan strongly recommends not rolling your own identity system unless absolutely necessary.
  • The risks include:
    • Choosing the wrong hashing algorithms
    • Falling behind on security updates
    • Missing modern requirements like passkeys
  • Identity is presented as a classic “buy, don’t build” category.

Boring technology still wins

  • The episode makes the case that “boring” technologies often win because they are stable, well-understood, and widely supported.
  • That applies to:
    • Identity systems
    • Password hashing
    • Even programming languages with long histories and large training corpora
  • The more standardized and boring the underlying system, the easier it is for agents and humans to work safely with it.

Practical Takeaways

For teams building with AI agents

  • Give agents their own identities where possible.
  • Use least-privilege access by default.
  • Sandbox agent behavior and constrain network/runtime access.
  • Prefer established standards like OAuth over custom auth logic.
  • Avoid collapsing human and agent identities except in narrow, well-controlled cases.

For software leaders

  • Measure AI effectiveness with business and engineering outcomes, not just output volume.
  • Expect security, operations, and scalability work to become more important as AI makes software cheaper to produce.
  • Treat AI as a productivity tool, not a magic replacement for engineering judgment.

Closing Thought

The episode’s core message is that AI agents do not eliminate the need for identity, authorization, or engineering discipline — they make all three more important. OAuth, least privilege, and careful system design remain essential as software moves toward more autonomous agent-driven workflows.