AI Software Development
AI agent development
Task-scoped agents with human-in-the-loop checkpoints, tool permissions, and logged decisions
- Duration
- 6 to 10 weeks
- Ladder stage
- Product
- NATIVE stages
- Integrate, Validate
The situation
When this engagement is the right one
An agent is a system that takes actions. The engineering question is not autonomy, it is permission: which tools, on which records, under whose authority, with what recorded afterwards.
You are probably seeing
- An agent demo impressed the board and nobody can say what it is allowed to touch.
- Multi-step chains fail silently and the failure surfaces as a customer complaint.
- Cost per task is unknown because nobody instrumented the loop.
What we do
The work, in the order it happens
- 01Scope the agent to a task with a bounded action space.
- 02Define tool permissions and the records each tool may read or write.
- 03Place human checkpoints where the cost of an error exceeds the cost of a review.
- 04Log every decision with inputs, tool calls, and the outcome.
- 05Set cost and loop-count ceilings with automatic termination.
What you get, and keep
- A production agent with a documented action space and permission model.
- Checkpoint design with the reviewing role named.
- A decision log and replay capability.
- Cost-per-task instrumentation and ceilings.
Prerequisites
- A workflow where the exception path is already understood.
- A named owner who accepts accountability for agent actions.
Not included
- We do not build open-ended autonomous agents with unbounded tool access.
- We do not deploy agents against production write paths without a checkpoint.
Duration and price shape
6 to 10 weeks
From a published figure, scoped by action space and integration count.
Where this sits in the method
Commercially this is a Product engagement on the Proof, Product, Platform ladder.
FAQ
Questions we get asked
- How autonomous should an agent be?
- As autonomous as the cost of its worst plausible error allows. That is an arithmetic question, and we do the arithmetic with you.
- What about multi-agent systems?
- Rarely justified in enterprise workflows at this stage. We will say so if your case is one of the exceptions.
Related engagements
Work that usually sits either side of this
AI Software Development
Custom AI Application Development
Production applications with authentication, data boundaries, audit trails, and an evaluation harness
AI Software Development
Enterprise RAG and Knowledge Systems
Retrieval architectures designed to survive audit: source attribution, permission inheritance, and evaluation
Next step
Ready to scope ai agent development?
Bring the pilots you already have running. The first call is a scoping conversation, not a pitch.