Inspectable Agent Systems
How agentic software should represent memory, authority, state, provenance, and evaluation.
Research Question
What architectures make agent memory, context, execution, evaluation, provenance, and authority boundaries observable and controllable without sacrificing adaptability?
Description
The lab's current investigation. Three of the five daily research desks independently converged on the same underlying problem: as agents take on more autonomous work, memory, context, execution, evaluation, and authority all need to become explicit and inspectable — not implicit behavior buried inside a model. This project tracks the graph-native memory architectures, independent-evaluation harnesses, and MCP/authorization patterns that are emerging as the answer.
Latest Lab Note
Today's lab state converged around inspectable agent systems: graph-native context, governed memory, independent evaluation, explicit authority boundaries, and simulation-based testing. The immediate study path is to compare procedural graphs and context architectures, then connect those findings to build opportunities in MCP, spatial intelligence, and agent evaluation.
Related Experiments
ProposedProcedural Graph Runtime
Can a validation-gated, self-evolving execution graph improve long-horizon agent workflows without increasing unsafe tool use?
hub Inspectable Agent Systems
ProposedIndependent Test Agent
Does separating an independent Test agent from a Repair agent reduce false-confidence patches without hurting resolution rate?
hub Inspectable Agent Systems
ProposedRevocable Memory Graph
Can a graph-based retriever enforce memory revocation structurally while still preserving useful personalized context?
hub Inspectable Agent Systems
ProposedSafety Judge Stress Lab
Which content-invariant wrapper transformations produce the highest disagreement across automated safety judges, and can simple ensembles resist them?
hub Inspectable Agent Systems
ProposedContext Graph vs. Context Database
Does explicit graph structure improve retrieval traceability and multi-hop reasoning over a context-database approach, at the same token cost?
hub Inspectable Agent SystemsRelated Open Source
Queued to studysemantica-agi/semantica
Graph-native context, provenance, ontology, and accountable-AI layer for agents — the #1 repo out of the Sept 9 run.
hub Inspectable Agent Systems
Queued to studyvolcengine/OpenViking
A live context database unifying memory, RAG, and skills — a concrete agent-memory architecture to test against Semantica.
hub Inspectable Agent Systems
Queued to studybrowser-use/browser-use-pi
Four days old and already worth dissecting: a minimal, evaluation-driven browser agent — small enough to fully understand.
hub Inspectable Agent SystemsRelated Sources
Procedural Graphs: Self-Evolving Execution Structures for LLM Agents
A localized, editable execution graph that biases the next action and can self-evolve — the strongest architecture fit of the Sept 9 research run (98/100 fit).
ExecCritic: Learn to Test, Test to Improve for Coding Agents
Independent Test/Repair agents in a fail-closed harness reach 72.6% on SWE-bench Verified, +11.4 points over a no-test baseline.
MeClear: Cooperative Game-Theoretic Attribution and Risk-Aware Memory Clearance
Reports 85.9% target recall and 82.3% task recovery across long-dialogue pools when clearing harmful/redundant agent memory.
WRITER: Building Agent Memory (engineering blog)
Task-, requirement-, and deliverable-oriented enterprise agent memory — the #1 California signal on Sept 9 (99/100), and the seed of the Context Graph vs. Context Database experiment.
LangChain: Organizing Context in a Multi-Agent Harness
Forked subagents inheriting supervisor context, compared against retrieval and graph-context strategies for reducing repeated lookups.
Arcade: Skills Over MCP
Runtime divergence around MCP is appearing before the spec fully lands — a live standards/governance conversation, not a settled protocol.