Copying between systems
Invoices, orders and status updates arrive by email, spreadsheet and partner feed. Someone retypes them into the system of record, every day.
Define, build, run and govern autonomous workflows, worked by AI agents and approved by your people. Everything on the left is an agent's proposal; nothing reaches the right until a person approves it.
An improvement moving through the loop: purple while the agents work on it, gold at your decision, green once it runs, red if you send it back.
Continuous improvement, on an infinite loop
Self-improving, under oath: every claim checked, every change on the record, every step approved. What that means
Human in the loop, where it matters. On your choice of AI provider.
AI agents do the routine work inside each workflow, from matching invoices to answering tenders. When a decision needs a person, it reaches the right one, with the evidence, the reason and the way back.
One decision needs you. Your agents handled the rest overnight.
Approve a $12,400 payment to Red Rock Freight? It is above the billing team's $10,000 limit, so it comes to you as controller. Pallas matched it to load LD-8851.
Copying between systems
Invoices, orders and status updates arrive by email, spreadsheet and partner feed. Someone retypes them into the system of record, every day.
Chasing approvals
Exceptions sit in inboxes. Nobody is sure who is allowed to approve, so work waits until someone notices it is late.
AI you can't hand the keys to
AI tools can do the work, but you can't see what they did or why, and you can't stop them from approving their own mistakes.
AOS Platform is the set of products every other product is built with. Each one has its own version, its own risk class and its own dial for how far agents may act in it.
Describe the work and drop in what you have. Agents write the brief and the definition, with its checks, and you read every rule back in plain words.
Each definition becomes a signed version, checked and published. The catalog shows what a version is built on, what it needs and what it works with.
Installs, sign-in, screens and every command run on the same engine, with agents doing the routine work and your people taking the exceptions.
The approval law, every crossing where a decision is needed, and the record of each one, sealed and witnessed outside AOS.
Your AI workforce: each agent sits in a declared seat, under a grant with a ceiling and an end date, at a level it has earned.
The loop: observation, proposal, preview, crossing, rollout, verification. It runs at every level, and nothing lands without a decision.
A publisher defines a platform, its products and their apps as signed models. You install them, fill them with your own people, agents, keys and data, and may add a layer of your own. Nobody can change what a publisher signed, and no publisher can see inside your company.
From one publisher, with a shared core: the records, roles and agent seats its products have in common.
A signed model with its own version, risk class and dial. Freight Ops is a product. So is AOS Define.
One main record, such as a load or an invoice, with its screens, its workflows and everything that acts on it.
A field, a step, a screen, a whole app. Signed under your own policy, versioned and rolled back like anything else.
Publish
A publisher defines and signs. AOS8 publishes AOS Platform. Flowwolf publishes Flowwolf AOS, which is built on it. Other companies can be admitted as publishers, and a publisher sees its installers as counts, never as records.
Install
Choose from the catalog, and AOS drafts the install for you: the plan, and the roles and seats filled from your own roster. Preview how it fits your records and settings, confirm each line, decide, and go Live. Pin a version or follow the publisher's. An upgrade is the same steps, and work in flight finishes on the version it started on.
Make it yours
Extend what you installed with a layer of your own, or define your own products beside it. A layer can tighten a guard and never loosen one, and before any upgrade it is replayed against the new version so you see what no longer fits.
People never have to learn the hierarchy to use it. They work in an app, and the platform and the product show as where things come from. The model, for developers →
A platform, a product, an app and a workflow all read the same way: six faces, one for each product of AOS Platform. Define shows what it is, Build how it is made, Run what it is doing, Govern who decides, Improve how it gets better, and Force who works in it. Learn one page and you can read every level.
set the dial of Load lifecycle to L1 let Iris verify "certify" in Partner onboarding switch off Tenders # refused, with the reason, if the app is required add a number field "dock door" to Loads add a step "second look" after "plan" in Load lifecycle make "tender" a crossing in Load lifecycle # the first three are settings; the last three change the model
No developers and no months-long project. The people who know the work describe it, and the people who are accountable approve it.
Say what you need and drop in what you already have: the spreadsheet, the policy, last quarter's emails.
Agents draft the definition and review each other's work. You try the version itself on a twin of your work: walk it as any role, replay it on past cases, or run it alongside today's work with no effect. What you preview is what you sign.
The signed version runs every day. When the agents spot a better way, they propose it with the numbers behind it, you preview it, and you decide.
Six AOS agents define, build, review and improve your products. Your own agents do the daily work, trained on how your team works and promoted only when they've earned it. Each one has a seat, a level and a report card, and each has something it never does.
Proposes the plan for each version, reads the other agents' work and reports.
Never decides, and never approves work it planned.
Works out the brief with you. Asks one question at a time and turns your documents into a complete brief.
Never drafts or reviews the design it briefed.
Drafts the definition from the approved brief: the screens, the rules, the steps and the tests.
Never reviews its own draft.
Reviews the draft, ideally on a different AI provider, and groups what it finds by cause.
Never drafts.
Watches how your processes actually run and reports what it saw, with the numbers.
Never proposes the fix itself.
Writes the signed code behind what a definition can't express, such as parsing a file. Tests first, then code.
Never reviews or approves its own code.
You don't build them from scratch. A product declares its seats; you seat an agent in one with a grant, such as matching invoices up to a limit until the end of the quarter. It starts by suggesting, and takes on more only when its record earns it and a person agrees.
Example report card. Figures are illustrative.
AOS improves itself in a loop. The same loop that improves your products also improves the agents that run them, and AOS itself.
Every improvement is previewed on your past work, comes with a rehearsed way back, and waits at a crossing for a decision. Done means verified against its target after a watch window. If it misses, it is rolled back. See the loop.
Steps that slow your team down get proposed fixes.
“14 of 40 invoices waited over 2 hours on a person.”Agents earn a higher level and sharper playbooks as their record improves.
“Let the matching agent check routine invoices under $10,000.”AOS is built by the same process it gives you: the same stages, roles, records and promises. If the process is good enough to build AOS, it is good enough for your product.
“Definer asks fewer, better questions after the last update.”Managers spend a few hours reviewing. The agents do the drafting.
One briefing across every product you run. Everyone sees only what waits on them, with a deadline and a reason.
Every item says when it is due and who it goes to if it is missed.
Who did what, why, and on whose approval. Ready for audit.
The first product sets AOS up for your business. Every one after that reuses your people, roles, rules and connections.
Match invoices to loads, orders or receipts, flag differences, and route payments for approval.
Build invoices from completed work, send them, and apply incoming payments to the right accounts.
Read orders from emails and attachments, create them in your system, and ask when something is missing.
Receive tenders, status updates and invoices from trading partners, and answer them on time.
Open a case, gather the evidence, ask the other party, and close it when the answer arrives.
Spend, credit and pricing approvals that go to the right person by amount, with cover when someone is away.
Flowwolf is a publisher. Its platform, Flowwolf AOS, is built on AOS Platform and adds what freight needs: products for brokers, carriers and shippers, and agents that speak your language of loads, lanes, rate confirmations, PODs and the partners you trade with.
A freight company installs both, and can add products of its own on top. It works alongside your TMS, accounting and carrier-pay systems, and carrier payments always need a person.
These are rules the engine enforces on every command. Your company can make them stricter. No setting, no publisher and no layer can make them weaker.
Every agent's work is approved by a person or, for routine items within your limits, by two different agents that have each earned their vote on that kind of work.
Agents start by watching and suggesting. They take on more only after they have proven themselves, and only when you agree.
Who did what, when, and on whose approval, in a record that can't be quietly changed and can be checked outside AOS.
Every change comes with a rehearsed way to reverse it. If an improvement misses its target, it is rolled back.
Passwords and keys are never shown. Personal data reaches an AI provider only if your policy allows it. Export everything, any time.
Every company runs on the same engine under the same rules, AOS8 and every publisher included.
Access controls stop an agent from reaching the wrong system. Decision controls stop it from approving the wrong payment. How AOS governs decisions →
Each agent has a report card: how often its work was right, how often a person changed it. When the numbers support it, the agent proposes taking on more, and a person decides.
How far an agent may act in a product is the lowest of four things: the level it has earned, the ceiling on its grant, the dial on the product, and the dial on the workflow if you set that lower.
Moving money, refunds and signatures stay with people at every level.
Your company lives in a cell, walled off from every other company in it. A cell runs one of two ways, and the rules, the record and every product are the same in both.
A product published once installs in a cell of either kind. Both kinds are designed to produce the same record, so a company can move from one to the other and be verified. The same design has AOS propose where your work should run, from what it measured, through the same loop, with the decision yours. The design, for developers →
Bring your existing records in and keep your current systems for what you choose not to move.
Use your company's identity provider, with roles mapped from your groups.
Data is encrypted at rest and in transit, with each company's data under its own key.
Each product can be offered to an outside AI client as tools. The client is seated like an agent, and nothing is open until you seat one.
Export all your records and their history, and verify them without us.
AOS8 is a Go core on Temporal and PostgreSQL with a React app. The rules that govern agents live in a small, tested core, and the record they leave can be verified without us.
# every change is an event, appended in the same transaction
event = [seq, tenant, type, entity_id, actor,
payload, commitment, prev_hash, ts, ...]
hash = sha256(canonical_json(event))
anchor = [seq, tenant, from_seq, to_seq, root,
prev_hash, ts, signature, key_id, ...]AOS8 is the engine beneath everything, and the organisation that builds it. It publishes AOS Platform: six products (Define, Build, Run, Govern, Improve and Force) for making and running autonomous work in any industry.
Flowwolf is a logistics publisher. Its platform, Flowwolf AOS, is built on AOS Platform and adds the freight products. A freight company installs both. Other publishers can build a platform for their own industry the same way.
No. The people who know the process describe it, and AI agents draft it. Your team reviews every rule in plain words, and previews the version on your own work, before anything runs. Developers are only needed if you want to connect a system AOS doesn't already reach.
Yes, by adding a layer of your own on top of it: a field, a screen, a step after an event, a whole app, or a product of your own beside it. You never edit what the publisher signed, which is what keeps it upgradeable. Changing the publisher's own steps is possible only at points the publisher declared. A layer can tighten a guard and never loosen one.
Agents propose; they don't act alone on anything that matters. Each proposal shows what the agent looked at and how confident it is, and the person or the agents reviewing it can change it. Every change has a rehearsed way back, and you can pause all agents in one click.
Two rules. An agent never approves its own work. And a verifier knows the work: every decision names a verifying seat, and an agent's vote counts only from that seat, once it has earned it on that workflow by voting in preview and matching what your people decided.
Two votes decide only when the agents differ in substance: no shared playbook and, by default, not the same AI provider. Otherwise a person decides.
Say it in a sentence, such as “add a step for a second look after planning”. AOS turns it into a plan, an agent drafts the change, and it waits for a person your policy names. A setting lands when it is signed; a change to a product lands when it is built. Either way it is on your record, with a way back.
The providers you choose. You can set one for the company, for a platform, for a product or for a single agent, and the most specific setting wins. A reviewing agent can use a different provider from the one that did the work. If a provider goes down, your processes keep running; only new suggestions wait.
Yes. Each product can be offered to an outside AI client as tools. The client is seated like any agent: every call is a command under a grant, what it may read is bounded by your data policy, and its reads are recorded. Nothing is open until you seat one.
In a cell that holds many companies, each walled from the others. Today a cell is Always on: processes that always run, on a database the cell runs itself. A second kind, On demand, is proposed and in design: it runs work when it arrives on a cloud vendor's services. The rules and the record are the same in both, and a very large or regulated company can have a cell of its own.
Each company's data is kept separate and encrypted under its own key. Your data policy decides which information may be sent to an AI provider at all. A publisher sees its installers as counts, never as records.
Your first process is usually defined in a working day of agent time, with a few hours of review from the managers who approve it. The first one takes the most effort because it sets up your people, roles and connections. Each one after that is much faster.
A witness under oath has to tell the truth, on the record, and answers for what they say. AOS8 improves itself the same way. When an agent argues for a change, every figure it uses is checked against the record before it's shown as fact. Every proposal, decision and change is written to a record that's witnessed outside AOS8, so anyone can check it later. The six promises always hold, no improvement lands until a person approves it, and if it misses its target it's rolled back.
“Under oath” is a standard we hold the software to, not a legal guarantee.
Most agent platforms help you build agents and control what they can access. AOS starts from the work: you sign a definition of how it runs, AI agents do the work inside it, and the engine decides who must approve each decision, by amount and risk. Every change can be previewed first and comes with a rehearsed way back, and the whole record can be verified outside AOS.
Yes. A product declares its seats, and you seat an agent in one with a grant. It learns from how your team does the work, starts by suggesting, earns more trust as its record improves, and is promoted only when a person agrees. You never write an agent from scratch.
AOS8 is built by Flowwolf. Flowwolf AOS, Flowwolf's autonomous back office for freight companies (brokers, carriers and shippers), is built on AOS Platform, the same platform you see here. AOS Platform itself works for any operation that can be described as roles, records and approvals.
Bring one process you'd like off your team's plate. We'll define it with you and show you what changes.