At QCon AI NY 2025, Tracy Bannon presented a talk examining how the rapid adoption of AI agents is reshaping software systems, and why organizations risk repeating familiar architectural failures if they treat all “AI” or “agents” as interchangeable.
Bannon argued that much of the current confusion stems from collapsing very different behaviors and risk profiles under the same labels. Bots were described as scripted responders that react to predefined triggers, while assistants collaborate with humans and remain largely under human control. Agents, by contrast, are goal-driven actors capable of making decisions and taking actions across systems.
Everyone is talking about AI ‘productivity.’ Very few are talking about the architectural amnesia that comes with it. – Tracy Bannon
To reify, Bannon outlined a set of autonomy patterns that commonly appear across the software development lifecycle. These ranged from AI-assisted tools embedded in existing workflows, through task-level agents that operate within bounded scopes, to multi-agent orchestration that coordinates end-to-end flows, and finally to mission-level autonomy where systems plan, optimize, and adapt toward higher-level goals.
A central theme of the talk was that autonomy does not fail on its own; failures occur when autonomy grows faster than architectural discipline. Bannon described this gap as producing what she called “agentic debt”. She connected agentic debt to familiar problem areas such as identity and permissions sprawl, insufficient segmentation and containment, missing lineage and observability, and weak validation and safety checks.
Bannon tied this risk to broader industry trends, noting research indicating that a large majority of technology decision-makers expect technical debt severity to rise in the near term due to AI-driven complexity. She argued AI does not introduce fundamentally new failure modes, but it magnifies existing ones by accelerating change and increasing the blast radius of mistakes.
She focused on applying established architectural principles to agentic systems. She argued that organizations already know how to manage risk in distributed systems, but often forget these lessons under pressure to move faster. Governance, in this context, was presented as the minimum set of controls needed to build trust, including clear accountability and traceability of actions and data flows.
Identity was highlighted as the foundational control on which other safeguards depend. Bannon stated that every agent must have a unique, revocable identity, and that organizations should be able to answer basic questions quickly when something goes wrong: what the agent can access, what actions it has taken, and how it can be stopped. She described a minimal identity pattern consisting of an agent registry.
We chase visible activity metrics … and quietly starve the work that keeps systems healthy: design, refactoring, validation, threat modeling. – Tracy Bannon
Decision-making discipline was another recurring theme. Bannon encouraged teams to start with “why” rather than “how,” and to make tradeoffs explicit before increasing autonomy. She described decisions as optimizations that always improve one dimension at the expense of another, such as value versus effort or speed versus quality.
The talk concluded with a call for architects and senior engineers to take an active role in shaping how AI agents are introduced. Bannon framed this responsibility as preventing architectural amnesia by designing governed agents rather than ad hoc automations, making risk and debt visible, and pursuing higher levels of autonomy only where they clearly deliver value. Her closing message was that the core practices of software architecture remain valid, and that the challenge is not learning entirely new disciplines.
Developers wanting to learn more can explore additional QCon AI sessions and InfoQ coverage, with recorded videos from the conference expected to be available starting January 15, 2026.
