AgentFlow
AgentFlow is the agent harness we build on. It connects any AI model to your systems and documents, adds memory and a team of specialist agents, and wraps it all in rules, approvals and cost controls. It runs in your cloud, and it’s yours to own.
Why a harness
Before it can take an action, it needs tools, memory, rules, approvals and a record of what it spent. AgentFlow is that layer.
The model on its own
The model inside AgentFlow
Whichever model you call this quarter will be a different one next quarter. The harness stays, and you own the code.
New to agents? Eight words to know first.
A large language model such as GPT, Claude or Gemini. On its own, it can only read text in and write text out.
Everything the model reads before it answers: who it is, the rules, background information and your question.
Roughly three-quarters of a word. Models charge per token and can only read so many at once.
Like a desk: only so much fits. Everything the model considers for one answer has to fit on it.
Instead of only answering, an agent can look things up, take steps and decide what to do next until the job is done.
Search the CRM, query a database, draft an email. Without tools, an agent can only talk.
The main agent passes parts of a job to narrower specialists, like a manager.
Tools, memory, rules, teamwork and tracking: what lets a model do real work. AgentFlow is a harness.
See it work
AgentFlow is the infrastructure around the model: tools, memory, approvals, evals and cost. It installs on AWS, GCP, Azure or your own hardware, and you can fork it.
Ships three ways: custom agents, in-app copilots and drop-in AI primitives.
How it works
Five steps, the same every time, whether the request starts in chat, an API call, a Slack thread or a schedule. Steps 3 and 4 repeat until the agent has what it needs.
Type in a chat window, or another app sends the request, or a button or schedule starts it.
The conversation so far, attached files, your preferences and notes, and relevant company documents.
Answer directly, use a tool such as a CRM search, or hand the job to a specialist agent.
Tools and specialists run one after another or side by side. Sensitive steps wait for your approval.
The answer appears as it’s written. Reports, tables and drafts open beside the chat.
Inside AgentFlow
The blocks are modular. Use what a job needs, and replace or add parts without a rebuild.
One doorway to any AI model, measured on every call.
Instructions built from labeled, reusable pieces.
Checks its draft against the request before answering.
Talks to you directly, or works to another agent’s brief.
Searchable libraries of your documents and websites.
Playbooks opened only when the task calls for them.
Key facts, a searchable archive and a wiki.
Finds what you discussed with the agent before.
Tone, standing instructions and model, per user.
Type @ to point at exactly the record you mean.
Drop in files; the agent reads them, even long ones.
Actions in your systems, approved where it matters.
Specialists the main agent delegates to.
A step-by-step plan you can see and edit.
Reports, tables and drafts you keep, edit and act on.
Stores large outputs; the agent works from a summary.
Keeps the agent’s working memory from overflowing.
Automatic checks on what goes in and comes out.
Monthly caps, checked before every model call.
Every call recorded: model, agent, user, speed.
Spend by model, agent and user.
Thumbs up or down on the exact run.
One request, start to finish
“Prep me for my call with @Acme Corp tomorrow.”
You see one conversation and one brief. Behind it, ten building blocks did their part, and every step was recorded.
@Acme Corp selects the account record with that id.
It recalls that you like one-page briefs ending with action items.
The request passes the company’s automatic checks.
The main agent hands work to two specialists: Records and Research.
Records searches the CRM: open deals, recent emails, last meeting notes.
Research checks the renewal playbook; the call-prep skill sets the format.
Each step uses its chosen model; the draft is checked against your request.
A one-page brief opens beside the chat, ready to edit.
It offers to send the brief to your team. Nothing is sent until you approve.
The cost is logged against the monthly budget, and you rate the brief.
Production and ownership
Teams we work with often spend 6 to 12 months building the approvals, the trace and the cost controls before anyone uses the agent daily. AgentFlow is that layer. It installs once in your environment.
Each organization gets its own database, so no data crosses between customers.
Industry-standard sign-in for users and for other applications calling the agents.
Each delegation, tool call and result is recorded in order, with token use and cost.
Close the tab and the work keeps running; reopen and it’s all there. Cancel any run cleanly.
AWS, Azure, Google Cloud or on-premises. Reach it by API, Python or TypeScript, or embed its interface.
Swap the model, the search engine or any tool without touching the core. No license to rent.
Make it yours
Each deployment adds its own agents, tools and content alongside the core, so it stays bespoke and still gets every upgrade.
Imported tools, new artifact types designed in chat, knowledge bases, app connections
Your agents, tools, skills, artifact types, scheduled jobs and data tables
Identical for every deployment, upgraded centrally
In 30 seconds
AgentFlow connects any AI model to your systems and documents, adds memory and a team of specialists, and wraps that in rules, approvals and cost controls. It runs in your cloud. You can extend it, and you own the code.
ChatGPT answers from a chat window. AgentFlow is the code you install, connect to your systems, and build your own agents on.
It runs in your environment, each customer is walled off, and agents only see what the signed-in user can see.
Every call is measured. Once the monthly cap is hit, the next call does not run.
Sensitive actions need approval, answers are checked before they go out, and every step is recorded.
