CyberFreezeDev
Loading
AI agents that act inside your tools, with a person approving anything irreversible.
An agent is a model with a goal, allowed tools and a loop. The hard part is deciding what it may do alone.
A chatbot answers questions. An AI agent does work. It is a language model given a goal, a set of tools it is allowed to call (search the CRM, read an invoice, create a task, draft an email) and a loop: look at the situation, pick a tool, read the result, decide again, stop when the goal is met or a person is needed.
That loop is what makes AI agent development different from building a prompt. The hard parts are not the model. They are deciding which tools the agent may touch, what it may do alone and what it must hand to a person, how it recovers when an API returns an error, and how you see afterwards exactly what it did and why.
We build agents the same way we build them for our own products. In Fint an agent reads bills and bank lines and proposes expenses, deposits and draft invoices, but nothing reaches the books until a person confirms it in an Inbox. In Desk an AI assistant can be handed tasks inside the CRM and project tracker. Inside our own engineering work we run agent-talk, a mailbox that lets agents working in different projects message each other, which we write about in multi-agent systems.
If you are unsure whether you need an agent or a chatbot, start with chatbot vs AI agent, and if a fixed workflow might do, read AI agent vs workflow automation. The broader idea is explained under agentic AI.
Each one starts from a job a person on your team does today, step by step.
Bills and delivery notes matched to records, with entries prepared for a person.
Read incoming bills, purchase orders or delivery notes, match them to records, flag mismatches and prepare entries for a person to confirm. Pairs well with our OCR and document intelligence work.
Enquiries qualified, checked against the CRM, assigned to a rep and given a draft reply.
Qualify new enquiries from forms or WhatsApp, look up the contact in your CRM, check for duplicates, assign the lead to the right rep and draft the first reply. One anonymous engagement used n8n, the WhatsApp Business API, a CRM and Claude for exactly this.
Figures pulled from several systems, compared with last week and summarised with source links.
Pull figures from several systems on a schedule, compare them with last week, write a short summary and post it to email or Slack, with links back to the source rows.
Tickets and approvals sorted, routed with a stored reason and chased when they stall.
Watch a queue (support tickets, approvals, failed payments), sort each item, route it with a reason attached and chase the ones that stall. The routing reason is stored so a manager can audit it.
A planner splits the request and worker agents each own one system.
Where one agent would need too many tools, we split the work: a planner that breaks a request into tasks and workers that each own one system. They pass messages rather than sharing one huge prompt.
Every step, system and judgement call in the job becomes the agent's task list.
We sit with the person who does the work now and record every step, every system they open and every judgement call. This becomes the agent's task list and the first test cases.
Each action is a narrow tool: run alone, draft for approval, or blocked.
Each action becomes a narrow tool with typed inputs, for example create_draft_invoice(customer, items) rather than open database access. We decide per tool whether the agent acts alone, drafts for approval or is not allowed at all.
Built on 30 to 100 masked past cases, with failures kept as regression tests.
We collect 30 to 100 real past cases from your data (with personal details masked where needed) and build until the agent handles them. Failures are kept as regression tests.
The agent only drafts on live work while your team compares results.
The agent runs on live work but only drafts. Your team compares its drafts with what they actually did. We tune prompts, tools and limits from those differences.
A person approves each action; proven low-risk ones may later run alone.
The agent goes live with a person approving its actions. Low-risk actions that are approved unchanged for long enough can be moved to run alone. High-risk ones stay behind approval.
Every run logged with cost. You get the code, prompts, tests and runbook.
Every run is logged with inputs, tool calls, outputs and cost. You get the code, prompts, test set and a runbook, and we can stay on for support.
The agent proposes and a person confirms, the same pattern Fint uses.
The pattern from Fint: the agent proposes, a person confirms. It is slower than full autonomy and far cheaper than undoing a wrong payment or a message sent to the wrong customer.
Separate keys per agent, read-only where possible and no tool that deletes.
Separate API keys per agent, read-only wherever writing is not needed, and no tool that deletes. If the agent is fooled by a malicious email, the damage it can do is limited by what its keys allow.
Caps on steps, spend and messages; at a cap, the agent stops and asks.
A cap on steps per task, on model spend per day and on messages per hour. When a cap is hit the agent stops and asks.
Each action is stored with the model's reason, so any decision can be traced.
Each action is recorded with the reason the model gave. When someone asks why a lead went to a particular rep, the answer is one click away.
A workflow follows fixed branches you drew in advance. An agent decides which step to take next based on what it reads. We often combine them: n8n handles the fixed plumbing (triggers, retries, credentials) and an agent node handles the steps that need judgement.
Yes. For sensitive data we run open-weight models on hardware you control. Tool calling quality varies between models, so we test your actual tasks on the candidate models before committing. See private AI deployment.
Narrow tools, per tool permissions, spending and step limits, and human approval on anything consequential. The agent cannot call a tool it was not given, and the tools themselves check inputs before acting.
Anything with an API. We build integrations with Google Workspace, the WhatsApp Business API, Razorpay, Shopify, Tally and most CRMs. Where a system has no API we look at email parsing, exports or an MCP server wrapping a database view.
Every run is logged, and we keep a test set of real cases that is rerun whenever prompts, tools or the model version change. You see approval rates and edits made by your team, which shows where the agent still needs help.
Last updated