Writing
Notes on making autonomous AI agents verifiable — auditing, governance, memory, and self-evolving verification.
-
August 2026
Why Every AI Agent Needs an Audit Layer
Agents now execute code, call APIs, and spend money on their own — yet most ship with no way to answer "what did it do, and can I trust it?" What an audit layer is, why pre-execution beats post-hoc logging, and how to add one without changing agent code.
-
August 2026
What Is a Self-Evolving Verification Agent?
Hallucination is the reliability bottleneck for AI agents. How a verifier checks whether claims are grounded in their sources, returns a structured verdict, and improves itself with a process reward that resolves advantage collapse.
-
August 2026
How Do You Benchmark Tool-Use Agents for the Sim-to-Real Gap?
Agents ace clean benchmarks, then break on typos, timeouts, and duplicate tool names. How to measure that sim-to-real gap with 22 realistic perturbations — and why scaling the model doesn't close it.
-
August 2026
Do AI Agent Skills Stay Safe Across Languages?
Public agent skills differ by language — English skews technical, Chinese skews media/social/finance — and safety signals are uneven. Why cross-lingual obfuscation slips past single-language detectors.
-
August 2026
What Is Charter-Governed AI Orchestration?
Give a swarm of agents a goal and a credit card — will they follow the rules and stay solvent? How a YAML constitution, a CFO that enforces spend, and an append-only ledger give a multi-agent system verifiable fiscal discipline.
-
August 2026
How Do AI Agents Remember Long Conversations?
Bigger context windows don't solve long-term memory. How workload-adaptive cascade retrieval answers easy queries in ~0.9 ms and only escalates to dense retrieval when a query is actually hard.
-
August 2026
Do LLM Reasoning Models Need All Their Memory in HBM?
Long chain-of-thought floods scarce GPU HBM with KV cache — but the top ~40% of tokens carry ~80% of the importance. How a semantics-aware memory hierarchy tiers the cache without dropping a single token.
-
August 2026
How Do You Secure Multi-User AI Agent Networks?
When everyone has an agent acting on their behalf, harm travels between users through shared tasks. How an auditable sandbox measures cross-user attacks with utility and attack-success rate reported separately.