Switch Costs, Spec Evolutions, and Durable State

Research

The Case for Meta-Specification in Agent Design

We already have a ton of frameworks for building agents, but this paper digs into how to actually let them propose and refine their own specifications over time. It's a step toward systems that get better just by running, not just by waiting for the next model update.

Measuring the Cognitive Load of Dynamic Agent Switching

This one breaks down the hidden switch costs of agents jumping between tasks. If you’re building anything with multiple triggers or concurrent users, this gives you a framework for knowing when the LLM is actually confused versus just context-switching.

Evaluating Agent Persistence in Non-Linear Workflows

Standard benchmarks usually have a single 'correct' answer, but real work often requires iterative backtracking. This provides a way to measure how well your agent recovers from dead ends during long-horizon tasks.

Comparative Planning Heuristics in Small vs. Large Models

It’s easy to assume smaller models do less, but they often make different types of planning assumptions. This research compares those assumptions to GPT-4 class models, giving you concrete data on when to use a cheaper, faster model for structured tasks.

Bandwidth Analysis for Multi-Agent Reasoning Chains

In multi-agent setups, reasoning isn't just internal; it's part of the 'cost' of coordinating. This formalizes how agents allocate reasoning steps based on how much they trust their partners' inputs.

Tools

LangGraph Streamlines Backend State Management

Engineers are spending too much time gluing memory providers into their agents. This release reduces that boilerplate in LangGraph, letting you swap backend contexts without rewriting your core logic. Practical plumbing for developers.