Protocore
A protocol-first runtime for agents that must keep working after the first response.
Agent systems, measured in the runtime
Build agents around durable state, explicit contracts, and observable work.
Protocore is a proprietary runtime for long-running, tool-using agents. It keeps model behavior behind protocols, makes state and artifacts first-class, and streams work as it happens. This site publishes the evidence behind those choices—including methods, limitations, code, figures, and downloadable papers.
The runtime is part of the experiment
Agent quality cannot be reduced to the model’s final sentence. Tool policy, state transitions, recovery behavior, artifact validation, and streaming all change what the system can complete. Protocore treats those mechanisms as testable parts of a runtime rather than hidden application glue.
Protocol-first boundaries
Models, tools, sessions, memory, events, and execution backends meet through explicit contracts. Implementations can evolve without making the core depend on a deployment stack.
Durable work
Runs are organized around persisted state and inspectable artifacts. Long tasks can be resumed and evaluated from what they produced, not merely from a terminal status string.
Immediate observability
Text deltas, tool activity, and runtime events are forwarded while work is in progress. Operators and products see the same execution rather than reconstructing it after completion.
Reproducible evidence
Public results are dated experiments with definitions, sample sizes, sanitized inputs, limitations, and generated artifacts. Product descriptions are kept separate from measured claims.
Research program
Our current publication program studies two recurring failures in agent systems: producing complete artifacts under output constraints, and evaluating tool-using work without confusing infrastructure status with task success. A broader systems article connects those findings to Protocore’s architecture.
01
Output convergence
How runtime controls affect the completion of large artifacts when generation is capped or interrupted.
02
Artifact-grounded evaluation
How to separate runtime health, judge output, policy compliance, and objective artifact validity.
03
Stateful runtime architecture
How protocol boundaries, persistence, tools, and streaming compose into an agent runtime that remains inspectable.
A claim is only as useful as its provenance.
Every quantitative statement published here is expected to point back to a method, a dated configuration, and a public-safe artifact. The reproducibility standard describes what that means in practice.