Service · AI agent development
AI agents that work acrossapproved sources and tools
I build multi-step AI systems for tasks that need information from several places. The agent retrieves permitted context, uses agreed tools and brings consequential decisions to a responsible person.
Last updated
30-minute intro call on Calendly
Each tool call is permission-checked and logged.
01 / When this fits
When the next step depends on the findings
A task may need a playbook search, a matter-record check and a proposed response, with clear limits on each source and action.
Answering one question takes several steps across documents and systems.
Knowledge sits in playbooks, documents and people’s heads, and needs to be usable.
The task needs judgement about which source or tool to use next.
Reasoning and sources must be visible before anyone relies on the result.
Keep a simpler workflow where it fits
Use an agent when the next step depends on the findings: search, compare and choose an approved tool. If the steps are predetermined, a fixed workflow may be easier to review.
Explore workflow automation02 / What I build
Define the boundaries before the tool calls
The model is one part. The rest is permissions, state, retrieval and the moments a person takes over.
Tool-calling agents
Agents that plan steps and call approved tools, with state kept across long-running tasks.
RAG and knowledge systems
Retrieval over documents, playbooks and domain information, with sources shown.
Document intelligence
Structured outputs from contracts, filings and correspondence, ready for review.
Permission-aware access
An agent sees and does only what the requesting user is allowed to.
Production engineering
Background jobs, retries and state handling so agents behave reliably under real load.
Human escalation
Clear handoffs to a person when confidence is low or an action needs approval.
03 / Control
Useful answers. Clear limits.
Define the agent's access and handoff rules before it runs.
AI prepares
Search approved sources, compare findings and suggest the next step.
Person decides
Review important actions and resolve uncertain answers.
System records
Log sources, tool calls, permissions and reviewer decisions.
Explore AI and data governance for the permissions, processing boundaries and evidence behind these controls.
04 / How we work
From one task to a dependable agent
A scoped agent workflow, connected retrieval and tools, permissions, review controls and verification against agreed examples. Continued support and model changes are scoped separately.
- 01
Agent design
The task, tools, data sources and decision points, agreed before code.
- 02
Working agent
Built, connected to approved tools and tested against agreed scenarios.
- 03
Knowledge and controls
Approved sources, access permissions and review handoffs.
- 04
Documented handover
How it works, its limits and the next steps, so your team can run it.
Implementation details
Connections depend on available APIs, permissions and data quality. The exact tools follow the agreed scope.
Security testing covers prompt injection and permission checks within the agreed assessment scope.
Questions
Before we start
What is the difference between an agent and automation?
Automation follows known steps and uses AI to prepare the work. An agent decides which steps to take, based on what it finds, within the tools and permissions it is given.
Can an agent work with our internal documents?
Yes, within the access you approve. Retrieval runs over the agreed sources, and answers show where the information came from.
How do you stop an agent from doing something wrong?
It can only use the tools and data it is permitted to. Consequential actions wait for approval, and uncertain cases go to a person.
Which models and frameworks do you use?
The choice follows the task and your constraints. I commonly work with LangGraph, LLM APIs, RAG and tool calling on Node.js or Python and FastAPI.
Next step
Bring a task that spans documents and systems
We will identify the context it needs and the decisions that should stay with your team.
30-minute intro call, or email hkhantanoli@gmail.com(opens in new tab)
What to bring to the call
- The task or question to handle
- The documents and systems it needs
- Where a person must review
Please leave out passwords, access keys and client data.