Topos is where teams define their AI agents, set what each one is allowed to do, and watch the work as it happens. Other Latere products connect to the same agents and share the same identities, permissions, and audit trail.
incidents.jsonl. @spawn-1284b. Scope inherited, network:write denied. audit:run-1284:redact-3. Autonomous agents promise to take work off your hands. Today they usually take it into a black box. Topos is built to keep that work visible instead.
Tool calls, sub-agent spawns, redactions, and failures all enter the conversation log, each with a timestamp, an identity, and any artifacts attached. You can watch live or read it later.
Each agent runs inside a policy ring that sets what it can read, write, spend, and spawn before it starts. A child agent inherits a strict subset of its parent. An out-of-scope tool call returns denied rather than a silent result.
Add approval gates, redaction passes, or executive review wherever a decision matters. Agents pause until the sign-off lands. You get full autonomy where you trust it and full control where you do not.
Topos is the place where agents are defined: a name, a scope, a model, tools, a budget, and review gates. Other Latere products use those same agents with the same governance already built in.
Topos is built for the point where the question is no longer whether you can run an autonomous agent, but how many you run and on whose behalf. Six surfaces, one model of work.
Every delegation, spawn, artifact, and refusal lands in a single stream. Filter by run, by agent, or by kind, and step in mid-run when something looks off.
delegate, spawn, artifact, review, deny, or deliver.infra/db. Two reference the same root cause (connection-pool drift). Pinned to thread. @spawn-1284c (worker, read-only). Task: dedupe auth-related noise. github.read_repo blocked by policy at org/research. Partial result returned. Closing. A single run can spread across a dozen agents and three generations of sub-agents. The graph shows the lineage at a glance: who delegated to whom, who is still running, and who failed.
An agent's scope is a document, not a dialog box. You can read it, version it, diff it, and hand it to compliance. Inheritance is explicit: a child gets the parent's scope minus whatever the spawn removes.
−network:write and it can never grant that back to itself.Asynchronous messages between agents, and from agents to people. Pin, mute, or route them.
Every agent in operation, with its role, owner, scope, and current cost.
The full transcript of inter-agent work, with kind tags and artifact links inline.
The lineage of a run as a live graph. Click any node to scope to its sub-tree.
Scope as a plain-text document: versioned, diffable, and attributable.
Every spawn, denial, redaction, and delivery. Append-only, indexed for queries, and exportable.
Each agent has a name, a role, a model, a scope, a set of tools, and a budget. Everything else is operational state. You can read an agent before it runs and again after it returns.
AI executes. Humans decide.
Define an agent, give it a scope, and watch the work happen in plain sight.