PaxLabs Research publishes foundational work on agentic systems. It is open access and written for builders. Fourteen papers so far, each advancing a concrete claim with a measurement protocol.
Natural language instructions move through well-defined stages into typed, closed-vocabulary intermediate representations before execution. This converts intent from ephemeral prose into structured data that can be validated, debugged, and audited.
What memory systems must provide, including persistence, typed schema, audit trails, and content-addressed retrieval, plus why the context window alone fails every requirement for agents that operate over days rather than minutes.
Single-run scoring, missing intent fidelity, absent provenance: the measurement practices that make agent benchmarks unrepeatable, and the multi-dimensional evaluation framework that fixes them.
How meaning degrades as it crosses agent boundaries, along with the typed-message, shared-state, and provenance primitives that keep a swarm reasoning about the same thing.
Safety mechanisms consume capability. We formalize the tax, measure it across intervention classes, and show which architectures pay least for the same assurance.
An agent that calls a tool must know what the call changed. We define postcondition verification for tool use and show why most agent stacks cannot prove their own side effects.
The sliding attention buffer is a workspace, not a store. This paper formally separates recall, retention, and provenance, then examines the failure modes caused by conflating them.
Distilling a model into a smaller one moves weights; it rarely moves the architecture of competence. Where distillation preserves agency and where it silently drops it.
Task completion rate is a single bit. We propose multi-dimensional scoring across intent fidelity, cost, provenance, and recovery, then show how rankings invert when the dimensions are counted honestly.
How fluid model behavior condenses into stable, retrievable knowledge, plus the architectural conditions under which an agent's skills survive contact with new experience.
Injected instructions are not miscommunication; they are a security boundary violation. The paper favors architectural defenses such as typed channels and capability scoping over linguistic defenses that attackers simply talk around.
Benchmarks shape the behavior they measure. We document evaluation-aware adaptation in deployed agents and propose measurement designs that resist being gamed.
The minimal primitive set for reliable orchestration, with a proof that removing any one of the three collapses coordination into gossip.
Phase transitions in language model behavior: which capabilities arrive discontinuously with scale, which never arrive at all, and what that means for matching model scale to task.
We write for builders. Every paper makes a concrete claim, defines its terms formally, and proposes a measurement protocol. We do not publish architecture descriptions without evaluation criteria.
Every paper advances a falsifiable claim about agentic system design. We do not write surveys or taxonomy papers without arguments.
Key concepts are defined precisely enough to be implemented and tested. Ambiguity is the enemy of reliable systems.
Each paper specifies how its claims could be empirically validated. Theory without a path to measurement is speculation.
We write for engineers and system architects deploying agents in production, not for conference reviewers. Clarity over citation count.
All fourteen papers are open access. Correspondence, replications, and disagreements are welcome. A claim you cannot push back on is not worth publishing.