Agents
The crew that produces an estimate: what each one reads, what it hands on, and where it stops and asks
How a run is organised
6Read and recommend
Geometry opens the file and Process reads a verdict out of it. Neither commits to anything: between them they produce evidence and a recommendation.
Human checkpoint
The run stops between the two phases. Everything after it commits to a way of making the part, and an agent that costs a die nobody agreed to has made a decision that was not its to make. Confirming the process releases the rest of the crew.
Cost the committed route
Once a person confirms the manufacturing route, the rest of the crew prices it: the tool, the piece, the confidence in both, and the pack that goes to a supplier.
Phase 1 · read and recommend
Phase 1Runs on upload. Produces evidence, commits to nothing.
1. Geometry
Phase 1Reads the solid and scores how hard it is to make
Opens the STEP file and pulls the envelope, mass, wall range, cored features, draft and feature counts, then scores seven complexity factors against the process family the shape belongs to. Mass tells you what the metal costs and nothing else; complexity is what explains the other two thirds of the price.
- STEP AP214 solid
- Part master record
- Geometric signals
- Complexity index 0 to 100
- Band · simple to very complex
Where it stops. Reads geometry only. It does not read manufacturing intent, because no STEP file carries any.
2. Process
Phase 1Decides what the geometry says it should be made by
Matches the signals against three process signatures and returns a ranked verdict with a confidence figure and the evidence behind it: every row traceable to a line in the CAD read-out above it. Processes the geometry rules out are named as ruled out, with the reason, rather than quietly priced.
- Geometric signals
- Complexity index
- Recommended process
- Confidence %
- Evidence table
- Viability per process
Where it stops. Recommends; it does not commit. The route is confirmed by a person, and the rest of the crew waits for that.
Phase 2 · cost the committed route
Phase 2Released by confirming the process. Everything here is priced against that decision.
3. Cost
Phase 2Prices the metal, the conversion and the burden
Takes the billet or blank mass against the finished mass, prices both at the grade's rate and scrap value, adds the conversion route for the confirmed process, then applies overhead and margin in that order. Where a grade is changed on the form, the material lines recompute against the same basis.
- Confirmed process
- Material grade
- Annual volume
- Rate master
- Piece price
- Cost breakdown
- Variance to target
Where it stops. With no ERP connection it prices off the Rate Master defaults and says so, rather than presenting a defaulted rate as a paid price.
4. Tooling
Phase 2Sizes and costs the tool, then divides it across the programme
Derives the tool's size from the part: a pattern cut oversize for machining and shrinkage, a die block sized by the force it has to hold shut, then costs it as a tool shop would, line by line. The build cost is divided across the pieces the tool will actually make, and that division is the amortisation line in the breakdown.
- Confirmed process
- Part envelope
- Annual volume
- Production life
- Tool build cost
- Lead time
- Tool life
- Amortisation per piece
Where it stops. Costs one tool for one route. It does not decide whether the programme should buy it.
5. Confidence
Phase 2Separates what was read from what was assumed and what nobody has
Sorts every input into read, assumed or unknown, and turns that into an AACE 18R-97 estimate class with the accuracy band the class implies. The ± on the headline figure is not decoration: it is the width of what is not known, itemised here so it can be argued with.
- Every input the other agents used
- Fields marked unknown on the form
- Estimate class
- Accuracy band
- What would move it up a class
Where it stops. Reports the class honestly, including when a marked unknown demotes it. It will not narrow a band it cannot justify.
What the agents are, and are not
Each agent narrates work the tool is genuinely doing: reading the model, scoring complexity, matching process signatures, dividing a tool across a programme. Every figure an agent reports is the figure on the panel it fills, so the console and the workspace cannot disagree.
This build is a mockup. The rules are standard cost-engineering formulas, not a model trained on your data, and the run is paced rather than computed. When the historical catalogue is released, the same crew reports figures a trained model produces: the agents are the interface, not the estimate.
Systems are named honestly
An agent that cannot reach a system says so and names what it fell back to. A rate defaulted from the Rate Master is never presented as a price somebody paid.
A person confirms the route
The crew recommends a process and stops. Nothing downstream is costed until a human accepts or overrides that recommendation.