AI AGENTS
Agents that do real work
An agent with a bounded job and scoped tools: it gathers, checks, drafts and prepares, then hands the result to a named person before anything consequential changes.
AI agents & automation
AI that does real work, with people approving what matters.
Foundry 41 builds AI agents that do bounded, useful work in your systems, LLM features inside your product, and agent-assisted software delivery. We also help engineering teams introduce AI into their own delivery process without losing the ability to review and validate what ships.
Last updated
What we build
What kinds of AI work does Foundry 41 build?
Three kinds, all built so a person can see what the AI did and approve anything consequential.
LLM INTEGRATIONS
AI features inside your product
Extraction, drafting, classification, search and assistants built into your application, with evaluation on agreed cases and fallbacks when the model is unsure.
AGENT-ASSISTED DELIVERY
Agents in the software pipeline
Agents that turn a spec into a pull request, with AI code review, human approval and monitoring around them. Details below.
AI transformation consulting
How can your engineering team use AI without losing control of the work?
Start with one software-delivery path, not a company-wide transformation promise. We work with your engineers to identify where agents can help, what must remain legible to people, and how to check the result before and after release.
01 · MAP
Choose a bounded pilot
Walk through a real task from request to release. Identify repositories, data access, failure modes and the person accountable for the result.
02 · VALIDATE
Define what good looks like
Write acceptance criteria and representative evaluation cases. Agree automated checks, security review and the human decision points before an agent changes the workflow.
03 · OPERATE
Make the gates repeatable
Set a pull-request and release path, monitoring and rollback expectations, and a way for the team to review failures and adjust the process.
The deliverable is an agreed pilot scope and validation plan for your team's engineering process; implementation or coaching can be scoped separately. We agree scope and fee before paid work. If your goal is to replace a cross-system business workflow, we start instead with the paid Workflow Blueprint.
Agent-assisted delivery
How do AI agents help deliver software without skipping review?
We call it the Software Factory: AI agents working across software delivery without giving up review. Agents draft the change and open a pull request; every change then passes the gates a senior team would insist on: automated evaluation, AI and human code review, an explicit approval, a controlled release, and monitoring after it ships.
01 · Spec intake
A person defines the change
A written spec states the change and how it will be accepted. The agent works from the repository's own conventions, tests, and skills.
Gate: accepted spec02 · Build
An agent opens a pull request
The agent works on a branch in an isolated environment and opens a pull request with its reasoning attached. It never pushes to the main branch.
Gate: pull request, not a direct push03 · Risk gates
Checks run on every change
Tests, type checks, and workflow-specific evaluations run in CI; riskier changes need more review. Failures go back to the agent or to a person, not to production.
Gate: checks pass04 · Review
AI review, then a human decision
An AI reviewer flags likely correctness and security problems. A human engineer reads the diff and decides whether it merges.
Gate: human approval05 · Release
Ship through your pipeline
Approved changes release through the existing deployment path, with a known rollback.
Gate: controlled release06 · Monitor
Watch it after it ships
Logs, errors, and cost are watched after release. A regression becomes a new ticket and goes through the same gates.
Gate: monitored in production
Internal reference: Sky, Foundry 41's own company brain, opens pull requests in our repositories. For a client, the agreed pilot scope or Workflow Blueprint sets which repositories, gates, reviewers, and release paths apply.
Control
How do you keep AI agents from doing the wrong thing?
By limiting what they can reach, checking what they produce, and putting a person in front of every consequential step. The controls are designed in from the start, not added after something goes wrong.
Company brains and governed workflows- Bounded access
Scoped tools and credentials per job, read-only unless a write step is agreed.
- Evaluation before release
Agreed cases the agent or feature must pass before it goes live.
- Human approval
A named person approves anything that changes a system of record, a customer message, or production code.
- Monitoring after release
Errors, cost and quality are watched in production; regressions route to a person.
Questions
Frequently asked questions about AI engineering
What AI work does Foundry 41 do?
Three kinds of AI build work: bounded AI agents, LLM features inside your product, and agent-assisted software delivery with evaluation and human approval. We also offer scoped AI transformation consulting for engineering teams.
What is an AI agent that does real work?
An agent with a bounded job, scoped tools and credentials, and a named person who approves consequential steps. It prepares, checks, drafts or changes things in your systems and leaves a record of what it did.
What is agent-assisted software delivery?
Agents take a written spec, work on a branch and open a pull request. Each change then passes automated checks and risk gates, AI code review, a human engineer's approval, a controlled release, and monitoring in production.
Do you offer AI transformation consulting?
Yes, for engineering teams with a specific delivery process to improve. We scope a pilot, define acceptance cases and review gates with your team, then agree separately whether to help implement or coach the rollout. For a cross-system business workflow, the paid Workflow Blueprint is the starting point.
Do the agents merge their own code?
No. Agents open pull requests; a human engineer reads the diff and decides what merges.
Which AI models do you use?
Whichever fits the task and your agreements. You hold the provider accounts and pay the providers directly, so the choice stays yours.
How do you know the AI is doing the job?
It is evaluated on agreed cases before release and monitored after. Failures and low-confidence results route to a person rather than into production.
Bring one task or one workflow
Start with a fit call.
Tell us the work you want AI to take on, the current review and release process, and who checks the result today.