Brief in progress — review the script and demo before recording.
This video aims to shift how we think about software now that agents can write code really well and generate fairly complex integrations, logic, and interfaces on the fly. In many ways, we are moving toward self-evolving and self-healing software, with architectures that aim to balance flexibility and reliability.
- Runtime: what happens around the agentic loop: the conversation thread, configuration, state, native tooling, external tools, persistence, data structures, and more.
- Primitives: the concept of primitives in agentic architectures: building blocks that shape an environment in which agents can safely create and use capabilities such as integrations, packages, and schedules.
- Boundaries: prompt injection remains an open problem, so we need to balance non-deterministic behavior with deterministic logic, keep humans in the loop, limit access to resources, and sandbox agents. These measures do not eliminate the underlying problem. Pliny the Liberator’s work could be a useful example here.
- Durability: moving beyond a basic agent is not trivial. We should use simple examples to explain recovery from critical errors, context compaction, offloading tool results to files, and pausing execution.
- Trust: LLMs can make mistakes, be tricked or confused, and get stuck in loops. We need to show these common failure modes so students understand where agents can be given more freedom and where LLMs should not be used at all.
We may want to showcase scenarios such as:
- Breaking down the logic of an existing agent, perhaps Goose.
- Building a simple agent and presenting its core mechanics.
- Showing how harnesses such as Pi can be steered programmatically, including by other agents.