How agents work · os.construction

Agents in os.construction are AI workers that know construction: cost codes, commitments, retainage, AIA billing, RFIs. They do the repetitive work on your real project and accounting data, then hand it to a person to approve. Agents do the work. You sign off. This page explains how that works.

What an agent is

An agent is a focused worker with one job, such as coding AP invoices or assembling a pay app draft. Each agent has:

A role. What it is for, in plain language: “Code every invoice to job and cost code and match it to its commitment.”
A scope. Which projects and record types it works on, and what it can read. Like a person, an agent has permissions.
Triggers. What starts its work: a new invoice in the AP inbox, a new daily log, a nightly scan, the start of a billing period.
Outputs. What it produces: drafts, flags and proposals, never posted transactions.

It reads the real data, not a copy

The biggest difference between os.construction agents and an AI tool bolted onto an existing stack is where they read from. Agents work directly on the data core: the same commitments, budgets, RFIs and pay apps your team uses.

That matters because construction answers depend on context. Whether INV-4471 from Volt Electric ($48,200) should be paid depends on PO-118’s remaining commitment ($42,050), on whether CO #14 is approved, and on whether a lien waiver is on file. An agent reading a nightly export would be working from yesterday’s facts. An agent on the data core sees the commitment as it is now.

Propose, then approve

Every agent action follows the same loop:

Trigger

Something happens: an invoice arrives, a super logs a hidden condition, the billing period opens.
Work

The agent reads the relevant records and does the work: codes the invoice, drafts the RFI, reconciles the SOV.
Proposal

The agent produces a proposal with its reasoning and sources: “Exceeds PO-118 remaining commitment by $6,150. Suggest: hold pending CO #14. Sources: INV-4471.pdf, PO-118, CO #14 draft.”
Human review

A person with the right permission reviews it. They can approve, edit and approve, or reject with a reason.
Action and record

Only after approval does anything post, send or change. The proposal, the decision and the result are written to the audit trail.

Show your work

A proposal is only useful if it is fast to check. Every agent proposal is designed to include:

What it wants to do, in one line.
Why, with the numbers that drove the decision.
Sources, linked to the actual records and documents it used.
What changes if you approve: which records, which amounts.

If an agent is not confident (a blurry invoice, an ambiguous cost code, two possible commitments), it says so and asks instead of guessing.

Where agents help most

Agents are best at work that is high-volume, rules-heavy and tedious, where a person still needs to make the call:

Coding invoices and three-way matching them to commitments and receipts.
Turning field notes into draft RFIs with photos and spec references.
Assembling pay app drafts, reconciling approved COs to the SOV, flagging missing lien waivers.
Watching every job each night for fade, overruns and schedule slip.
Chasing expiring COIs and missing compliance documents.
Drafting owner reports and WIP commentary.

See The six launch agents for each one in detail.

What agents are not

Not autonomous. They do not act on their own authority.
Not a black box. Every proposal shows its sources and every action is recorded.
Not a replacement for judgment. A PM still decides whether a change is billable; a controller still signs off the WIP.
Not vague about your data. Your data stays yours. How models are hosted and how your data is handled will be shared with founding contractors before any agent runs on their jobs.