Blog
Research and observations from building agent software.
L0 to L5: where AI agents land on the autonomy spectrum
A survey of how AI agents are shifting from human-approved to human-supervised — autonomy levels, product positioning, and what it means for agent runtimes.
research
MiroFish: swarm prediction through social simulation
We examined MiroFish's architecture — from OASIS swarms to GraphRAG extraction. How it predicts the future and where the approach breaks down.
research
Async compaction: the race conditions nobody talks about
We examined how frameworks handle background compaction. Most avoid it entirely. The ones that try it have documented race conditions and silent data loss.
research
Built-in agents: what ships in the box
We surveyed ten AI coding products to catalog their built-in agents. Three have real multi-agent delegation. The rest ship modes.
research
Built-in tools: what your agent can reach
We cataloged the built-in tools in ten AI coding products. Some ship 5, some ship 20+. The differences reveal what each product thinks an agent should do.
research
Hermes memory: five layers, one learning loop
We examined Hermes Agent's five-layer memory system — procedural skills, Honcho user modeling, FTS5 search — and asked what it costs to remember this much.
research
Mem0: what three memory scopes actually cost
We examined Mem0's extraction pipeline, conflict resolution, and benchmark claims. Smart memory management is real — but most agents don't need three scopes.
research
Hermes Agent: what Nous Research built
We examined Hermes Agent's architecture — from Atropos RL training to persistent skill documents. Here's how it works and where it fits.
research
When agents share a channel
We tested what happens when two agents join the same group chat. Most frameworks have no built-in answer for the infinite loop that follows.
research
Context compaction in agent frameworks
We surveyed eight frameworks to understand who compacts what in multi-agent systems. The answer: each agent handles its own, and nobody coordinates.
research
Tool permissions and the bash bypass problem
If your agent has bash, do file-editing tools even matter for security? We surveyed eight frameworks and found the answer is architectural, not policy.
research
How agents call agents
We surveyed eight frameworks to map who can call whom in agent teams — upward, lateral, recursive. Most are more restricted than expected.
research
SOUL.md: brilliant idea, brittle implementation
We analyzed OpenClaw's SOUL.md personality system — what it gets right about agent identity, and why static files break under production pressure.
research
How AI frameworks control model thinking
We surveyed seven agent frameworks to understand who controls reasoning depth — the framework, the API, or nobody. The answers split three ways.
research
What should an agent capability bench test?
A survey of existing agent benchmarks and 120+ questions we think a practical capability bench should answer.
research
How developers configure their AI agents
A survey of CLAUDE.md adoption, MCP servers, skills, and hooks — and why your prompt system decays predictably from MVP to production.
research
Sandboxing AI agents: beyond Docker and WASM
How Claude Code, Cursor, Codex, and Devin gate bash access — and why Docker, WASM, and OS sandboxes are all solving the wrong layer.
research
Why multi-agent workflows fail in production
A survey of real coordination failures in Claude Code, Cursor, Devin, and the broader ecosystem — and what they reveal about agent architecture.
research
How AI agents remember: a survey of persistent memory
We surveyed how Claude Code, ChatGPT, Cursor, Windsurf, and others implement persistent agent memory — storage formats, compaction, and retrieval.
research
Plans vs tasks: how AI agents think before they act
A survey of plan-then-execute patterns across Claude Code, Cursor, Devin, Windsurf, and Copilot — and what it means for autonomous agent design.
research