Agentlas OS

Network 2.0 — staffing agents for a task

Finds and hires the right agent for a task from three sources — your local machine, your private Cloud, and the public Hub — without a hidden ranking algorithm making the final call.

Who decides

Your connected model reads the actual candidates and their qualifications, then writes the selection itself. The system doesn't silently pick a "best match" behind the scenes — it verifies governance, headcount, and integrity on the selection you made, and pins the exact release it prepares to run.

Scopes are sealed, not blended

local — only agents registered on this machine.
cloud — only your own private Agent Cloud assets.
hub — only public Hub specialists.
network — the sealed union of all three. A source that fails stays failed; it is never silently swapped for a different scope.

The WorkOrder — how a hiring request is shaped

Your connected model doesn't just fire off a search string. It writes a structured, de-identified WorkOrder describing the actual job before anything is searched.

taskBrief

What the work actually is, in redacted form — no private files, secrets, or raw local context leave the machine.

roleSlots

One or more role definitions to fill, each with its own required skills/knowledge and cardinality (how many hires that slot needs).

edges

How slots hand off to each other — the shape of the mini-team, not just a flat list.

forbiddenCommunities

Explicit exclusions the selection must respect.

selectionPolicy

Governs how the final pick gets made — see "Who decides" above.

The four steps, in order

search_candidates — federate Local + Cloud + Hub against the WorkOrder, returns real candidates with qualification evidence.
Your model authors a Selection from what it actually read — not a score threshold picking for you.
validate_selection — checks governance, privacy, identity, and headcount against the WorkOrder; does not re-rank or substitute your pick.
prepare_execution — pins the exact release, package hash, and content digest for each hire before anything runs, and re-verifies those digests to fail closed on drift.

This is not the same thing as a project's agent roster

A project keeps a small, persistent, human-curated list of agents (see Projects & staffing in Trust & safety). Network staffing is the opposite: a fresh WorkOrder, scoped to a single task, that can reach outside that list to hire a specialist for just that job.

/agentlas network split this launch into research, copy, QA, and release agents
/agentlas local   use only agents registered on this machine
/agentlas cloud   use my saved finance analyst agent
/agentlas hub     find public specialists for accessibility QA