research engineering

Agent systems need a memory control plane because retrieval safety and historical truth require enforceable policy and clear lineage, not mutable notes.

oss engineering

We built a compiler that turns natural-language metaphors into running programs. The metaphor is the source of truth; the code is the build artifact.

oss engineering

We built a tool that asks LLMs structurally impossible questions and classifies how they fail. The failures are more interesting than we expected.

oss engineering

An open source tool that breeds, mutates, and selects LLM prompts across generations.

oss engineering

We gave an agent 50 memories and let them decay. After 30 ticks, it retained 8 — exactly the ones it was still using.

engineering research

The previous Let Machines Talk posts described the theory. This is what it looks like as running code.

engineering research

Stopping the system is half the problem. Undoing what it already did is the other half.

engineering research

Observability tells you what's happening. Lineage tells you why.

engineering research

The question isn't whether your system will fail. It's how much it takes down when it does.

engineering research

A kill switch is a button. A kill path is an engineering discipline.

security opinion

Prompt injection is one attack shape. Authority theater, urgency pressure, and trust laundering across tool boundaries are the rest.

research engineering

Agent systems get designed for launch day. The first hour after failure is usually an afterthought.

research engineering

The term artificial intelligence gives teams an excuse to avoid accountability. These systems are real. Treat them that way.

research security

An agent that can produce a good plan is not the same as an agent that should be allowed to run it.

opinion research

Machine-to-machine communication isn't the risk. The risk is building the channels without building the controls.

research engineering

Before an agent acts, force its plan through a thousand bad futures. Ship what survives.

research engineering

What if an AI system never produced text at all — just causal graphs, state machines, and consistency proofs?

research engineering

Five directions for AI systems that aren't built on token prediction: deterministic reasoning, reversible execution, constrained testing, native substrates, and symbolic computation.

security identity

When people use machine credentials, intent is blurred and audit trails break. Agents make it worse.

security identity

Reused identities and weak isolation turn local incidents into systemic ones.

security identity

Leaked secrets become durable agent memory. That's why breaches linger.

oss engineering

EntropyOS is a Python runtime where APIs, code paths, and state have expiration dates built in.

security oss

An adversarial test exposed unsafe text propagation across agent boundaries. We found it, fixed it, and kept the test.

security identity

Over-scoped identities turn harmless tools into high-impact actions. The agent is rarely the root problem.

oss education

An open-source collection of text adventure games that teach security, ops, and incident response through interactive scenarios.

security identity

Agent incidents usually start with a mundane identity fault, not a novel model exploit.

news guides

A short list of news sites, blogs, and newsletters worth following for AI developments in 2026.