The Loop Is the Product
An agent is a language model inside a loop, and the loop is the part a team builds, measures, and stops. The model call is one node in it.
Large Language Model Agents.
Agents as systems: tools, memory, planning, safety
Notes
Sorted in orderReasoning Is a Scaffold
Chain-of-thought buys the model room to work, and self-consistency votes across sampled paths. Only a check outside the model can fail a wrong answer.
Tools Turn Guessing Into Looking
ReAct alternates thinking with tool calls so the agent gathers what it cannot guess. The action space decides how much guessing is left.
Frameworks Are Coordination Machines
A framework earns its keep when it makes a hidden decision explicit: who owns state, who calls tools, and what ends the loop.
Enterprise Agents Need Grounding
A larger context window holds more noise as easily as more truth. Grounding shows which shelf the answer came from, and enterprise work runs on that difference.
A Prompt You Cannot Score Is a Preference
DSPy replaces prompt tweaking with a signature and a metric. The metric you pick becomes the system you get, so design the metric before the pipeline.
Coding Agents Live or Die by Their Interface
Better commands and clearer observations improved SWE-agent results with the base model unchanged. The interface is the part you build, and SWE-bench still passes patches that miss the intent.
A Workflow Agent Needs a Tape
Models solve single steps and lose the thread once the steps compose. A replayable tape buys more reliability than a better prompt.
Language for Ambiguity, Solvers for Commitments
Language handles messy intent and a solver handles exact constraints. The translation between them is the step that breaks, so expose the formal problem before the solver commits.
Embodied Agents Need a Data Pyramid
A robot motion has no undo. An embodied agent needs a data pyramid, a simulator for cheap failure, and feedback that reports what changed.
Open Models Need Open Science
Two teams can run the same model and report different scores. The scaffold is part of the result, which makes disclosure part of the science.
Responsible Scaling Needs Measurement
Governance works when a measurement can change a launch decision. A gate needs an eval, a threshold, a required mitigation, and an owner who can refuse.
Untrusted Text Stays Evidence
Prompt injection is the native attack on any system that mixes instructions with untrusted text. The defense is architecture that keeps the two channels apart.