Agents that do the work
Each agent takes on one kind of recurring work: intake, reconciliation, review, triage. It carries the work start to finish, inside the systems you already run.
Plabs / Products / Intelligence Fleet
Autonomous operations Human-supervised Runs in your cloud
A supervised AI workforce for the work that never stops.
Intelligence Fleet puts a team of AI agents to work on your recurring, rule-bound operations. They handle what they can on their own, send every judgment call to your people, and only run when there is real work to do. You pay for work, not for waiting.
01 / What it is
The short version
A workforce you can trust with the busywork.
Your people stay in charge
Anything that needs a decision, or a step only a person should take, goes to your team. Agents work strictly inside the lane you give them, never around it.
Works with what you have
Agents act through your existing tools and data. If you already run Intelligence Fabric, the fleet plugs straight in; if not, we stand up the access they need.
02 / How it works
In plain terms
Work comes in. Agents do it. People make the calls.
Four moving parts, one clear flow.
- Work arrives. From your systems, on a schedule, or handed over by a person or another agent.
- An agent picks it up. The one that owns that kind of work claims it and does it end to end.
- It finishes, or hands off. Cleanly, through one shared queue. Nothing gets dropped, nothing gets done twice.
- Judgment calls go to your team. A decision to make, or an action only a person should take. The agent files it with the context and keeps moving; the fleet never stalls waiting on a human.
Each agent owns a domain: the slice of work it handles. We discover those domains from how your work actually flows, or your team defines them outright. Either way, the boundaries are explicit.
How work flows
- 01 / Queue
Work arrives and waits in one shared queue.
- 02 / Wake
A watcher spots it and wakes the right agent. While the queue is quiet, no model runs: idle means $0.
- 03 / Work
The agent claims the item, does it, commits. Then the next one, until the queue is dry.
- 04 / Resolve
Finished work lands as done. Judgment calls are filed to your team's review queue, recommendation attached.
- 05 / Loop
Your answer re-queues the item. Agents never block; the fleet keeps moving.
03 / Oversight
Why it's safe to run
Autonomy with a person's hand on it.
One place your team works
Decisions and human-action items land in a single review queue, each with its context and a recommended answer. No dashboards to babysit, no agent left hanging.
Agents touch only what they should
Every agent acts with its own scoped access to your tools and data. No shared keys, no standing back doors, no reaching past the work it was given.
Every action is on the record
Who or what did it, on whose behalf, and why. Logged for every step, in one trail your security and compliance teams can actually sign off on.
It runs where your data lives
The whole fleet runs in your cloud, on your accounts, inside your perimeter. Nothing is centralized with us; your data never leaves your control.
The economics
You pay for the work, and only the work.
Most "always-on" AI bills you around the clock, busy or not. Intelligence Fleet is the opposite. An agent with nothing to do isn't running, so it costs nothing. Cost tracks the work itself, and you scale throughput by adding agents, not by writing a bigger blank check.
An idle agent costs
$0
Nothing to do means nothing running. You're billed for busy periods, not for standing by.
Spend stays
Capped
Every agent works within a set budget per run. Need more throughput? Add agents, each with its own ceiling.
Coverage
24 / 7
Work that arrives at 3am is handled at 3am, or waits in your team's queue for the morning.
04 / Under the hood
For the architects
The machinery, for the people who'll operate it.
Everything above is the promise. Here is how it's built, the part your platform and security teams will want to take apart. Skip it if you just wanted the what; read on if you want the how.
An agent is a busy period, not a daemon.
Nothing runs while the queue is dry. A lightweight watcher checks for work on a short cycle (a cheap read, no model involved) and starts a fresh agent session only when there is something waiting. The session claims work, does it, commits, and re-checks until the queue is empty or its budget is spent. Then it stops. Idle agents run no model and hold no session, so cost and conversation history never accumulate between busy periods.
A judgment call doesn't block the run. The agent files it to your team's queue as a tagged item, either a decision (options and a recommendation attached) or a human action (the exact step to take), and moves on. Your answer re-queues the item and the work resumes.
# cheap poll: is there work? fleetctl scan → invoice-intake waiting: 3 start → vendor-recon waiting: 0 idle · $0 → access-review waiting: 1 start # agent session (work until done or budget) fleetctl run invoice-intake → start… budget 20m scope invoices:read,write → claim WQ-4711 → do → commit done → claim WQ-4712 → do → commit done → queue 3 → 2 → 1 → 0 ESCALATE WQ-4713 decision → your team's queue 2 options + recommendation attached STOP → session retired · idle · $0
CAP / 01SCHEDULING
Wake-on-work, not on a timer
A watcher reads the queue on a short cycle. No model runs, so the check costs effectively nothing. An agent launches only when there is pending work, a failed check to re-drive, or a scheduled sweep. Crashed sessions are caught by liveness and restarted. No cadence, no idle burn.
- Idle cost
- $0
- Poll
- No model
CAP / 02BUDGET
A budget on every run
Each agent carries a per-session budget and a deadline as configuration. The session keeps claiming work while work exists, context stays healthy, and it is under budget. Then it stops. Spend is bounded by design, and both budget and model are set per agent.
- Budget
- Per session
- Model
- Per agent
CAP / 03QUEUE
One governed work queue
Every unit of work is a durable record with an owner, an assignee, a risk class, and a full history. Handoffs between agents are atomic reassignments: no shared-file contention, no lost or duplicated work. A wrong route re-routes; nothing falls on the floor.
- Handoff
- Atomic
- State
- Durable records
CAP / 04FLEET
A mixed fleet, one registry
An assignee is a principal, and a principal is one of three kinds: an on-device model loop, an SDK-driven agent, or a person. The same queue routes each item to whoever handles it best. One registry, one audit trail, whether a machine or a human does the work.
- Kinds
- Loop · Agent · Human
- Interface
- One queue
CAP / 05DOMAINS
Domains, discovered or defined
Each agent owns a domain: the slice of work it is responsible for. We discover candidate domains from how your work actually flows, or your team defines them outright. Either way the boundary is explicit, and its escalation posture (what an agent may decide alone) is set with it.
- Source
- Discovered · defined
- Boundary
- Explicit
CAP / 06GOVERNANCE
Scope, audit & escalation
Agents act as scoped principals through your governed tools. Every action logs identity, scope, parameters, and a trace ID. Two non-blocking escalation classes, decision and human-action, route to your team on the risk classes, thresholds, and operations you configure.
- Identity
- Scoped principal
- Audit
- Append-only
05 / Engagement
How we work
Live inside a quarter, wider one agent at a time.
Map the work
We find the recurring, rule-bound work worth handing to agents, agree the domains and their boundaries, and set the escalation posture: what an agent may decide alone, and what always comes to a person.
~ 3–4 weeks
Stand up
The first two or three agents go live against your governed tools, the watcher starts handling real work, and the review queue and audit trail run end to end.
~ 6–10 weeks
Widen the fleet
New domains come online, budgets and models are tuned per agent, and oversight agents watch for drift. Throughput grows one agent at a time, never one big-bang cutover.
~ 6–12 weeks
Hand off
Your team runs the fleet and works the queue. Plabs stays as strategic advisor: quarterly reviews, new agent patterns, escalation when you want us.
Ongoing
06 / Specifications
For the architects
Straight answers to the questions your architects will ask.
- Deployment
- Runs in your cloud, on your accounts. Agents, watcher, queue, and audit all live inside your perimeter. Deploys to AWS, GCP, or Azure; on-prem available with enterprise terms.
- Tools
- Agents act through a governed tool suite over OpenAPI and MCP, the same surface Intelligence Fabric publishes. If you already run the Fabric, the fleet plugs into it; if not, we stand up the tools the agents need.
- Agent types
- An assignee is a principal of one of three kinds: an on-device model loop, an SDK-driven agent, or a human. All share one queue, one identity model, and one audit trail. Mix them per domain as the work demands.
- Identity
- Each agent acts as a scoped principal federated from your IdP. Tool calls carry that identity end to end; there are no shared agent credentials in flight.
- Audit
- Every agent action and every queue transition is logged with principal, scope, parameters, and trace ID. Append-only. Exportable to your SIEM.
- Escalation
- Two non-blocking classes: decision and human-action. You configure which risk classes, dollar thresholds, and operations always route to a person before an agent may act.
- Pricing
- Engagement-based, not per-seat. Map + stand-up + widen is a fixed scope; ongoing advisory is separately contracted. Agent run-cost is your own model spend; we do not mark up infrastructure.