Announcements / Notes / Research
Blog
On autonomous systems, human-in-the-loop AI, and tools that keep judgment visible where it matters.
The identity fabric: who acts, whose keys, who pays
One piece of agent work touches a runtime, a model gateway, a data store, and a command-line tool, and under every step sits a question: who is acting, whose credentials are spent, and who pays? The identity fabric roots everything an agent does in you, and seals each product's keys so a problem in one place stays in one place.
Drive: one home for your files and everything your agents create
Agents make real things, and today their output lands in your folders or vanishes with the session. What changes when your files and everything your agents create share one home, where your files stay yours and agent work is saved, kept separate, and easy to review?
AI as an Infrastructure
AI is becoming infrastructure: a layer other software depends on, with uptime, cost, observability, and trust all in play. This book follows the causal thread behind the modern AI stack, from compute and data to deployed agents.
Agent harness design: the cross-cutting contracts
The seams that cross primitive boundaries and belong to none: concurrency, operational failure modes, transport, observability, and evaluation. Where production agent platforms most often break.
Agent harness design: the governance fabric
An agent is a document: a versioned, owner-signed record of its prompt, tools, budget, and approval steps. The case that governance, declaring what an agent may do and whom it may delegate to, is a primitive distinct from identity.
Agent harness design: the execution runtime
The per-session execution loop: agent lifecycle, MCP connections, context management, scheduling, multi-agent composition, tool dispatch, and the interruption primitive that human approval depends on.
Agent harness design: the model-access fabric
Model access is not a config line on the harness. It is a separable primitive: who holds the provider key, which model answers, on whose budget, attributed to whom. A deep dive into the contract and where it fails.
Agent harness design: the identity fabric
The substrate every other primitive consumes: resolving each caller to a user, tenant, agent, and workload, with tenancy, delegation, isolation tiers, and the cascades that make right-to-erasure tractable.
Agent harness design: the persistence fabric
Durable state for agents: intent versus step logs, workspace and snapshot recovery, branching, artifacts, long-term memory, and the audit and erasure obligations that ride on all of it.
Agent harness design: the sandbox fabric
Where the agent runs, and what the isolation boundary actually buys. Trust-boundary placement, what isolation does not solve (prompt injection, overprivileged actions), non-bypassable egress, and supply-chain risk.
A reading, and nothing else
Somewhere there is a folder called "to read" that never empties: invoices, contracts, scanned forms nobody has time to open. Lectio reads one document and hands back exactly what you needed from it. One thing, done well, and nothing else in the way.
Topos: the place where agents work in plain sight
AI agents take work off your hands, and often out of sight. Who is running, what they are allowed to do, what they were refused: all of it tends to disappear. What changes when every agent has a name, clear limits, and a full record you can read?
Lux: model access for agents
Giving agents access to AI models sounds simple, until they start running in many places at once, each one needing keys. How do you give every agent the right model without handing your real provider keys to every process?
Adversarial verification: rethinking trust in multi-agent systems
More AI reviewers should mean more confidence, but if they share the same blind spots, agreement can hide the risk. Should we smooth over disagreement, or design the system so it has to reveal where an answer might be wrong?
Cella Sandbox: fast enough to throw away, durable enough to keep
Quick cloud functions forget everything; full machines cost money to keep running. What sits in between when an AI agent needs its own private space, a memory it can keep, and the freedom to vanish when it is idle?
Agent harness design: trade-off analysis
Every way of running agents trades speed for control and freedom for oversight. What has to be spelled out before an agent can safely touch your files, tools, accounts, and people?
Wallfacer: structure kills discovery, freedom causes collapse
Too many rules and the AI stops exploring; too few and the whole effort falls apart. After dozens of rounds and sixty thousand lines of code, what kind of workflow keeps agents creative without losing the thread?
Why Latere
As AI takes on more, the decisions that matter slip out of view, yet good results still depend on human judgment, limits, and taste. If the intelligence that matters becomes hidden, what should a company build around it?