Give the work
an organization.
Define who does what, who reports to whom and which tools each role can use. Alchemy expresses the team’s structure in an organization description, with reference teams you can adapt to your work.
01 / Why now
Everyone built agents. Few can manage, see or govern them.
The pilot worked. Then finance asked what it would cost at scale, and security asked who approved the actions. Nobody in the room could answer.
40%+of agentic AI projects are expected to be cancelled by the end of 2027, on cost, unclear value and weak risk controls.
21%of enterprises have a mature governance model for the AI agents they plan to run.
82%found AI agents in their environment in the past year that nobody knew were there.
Three numbers, one gap. Nobody could see what the agents were doing, nobody had decided what they were allowed to do, and nobody could prove what they did.
The fixThe fix is not a better agent. It is a way to manage, watch and govern every agent, with a person at the points that matter.
02 / The operating model
The agent loop is only the beginning.
A harness gives an agent the machinery to work: tools, context and an execution loop. Modern frameworks also offer memory, orchestration and human approval.
Those capabilities still need an operating model. Someone must define responsibilities, delegation boundaries, decision points and the record the work leaves behind.
ANATOMY OF AN ORGANIZATIONILLUSTRATIVE MODEL
ROLES → DECLARED ACTION → REVIEW
Team structure.
Decision points.
A record to inspect.
Defined roles. Connected responsibilities.
Give the work
Make the decision
points explicit.
Decide which supported actions require your approval. When a declared action reaches that boundary, the run waits. Inspect the proposed effect and supporting artifacts before deciding whether it proceeds.
Keep the context
behind the output.
A finished artifact leaves questions: which inputs were available, what happened along the way, and what needs review? Alchemy retains context bundles, events and artifacts on your infrastructure so you can inspect the work.
Inspect an illustrative record
EXAMPLE / DECLARED ACTION
- Role
- incident_commander
- Action
- Release the Day Three response pack
- Inputs
- Exercise setup, scenario knowledge base
- Artifacts
- response-suite.docx, precheck.md
- State
- Waiting for approval
An illustration of the information available for review; no product action is being executed.
Start with a team
you can run again.
Carry the organization forward to the next bounded goal. Adapt its roles, tools and rules as your work changes.
Alchemy brings team structure, declared controls and recorded work into one runtime.
Explore the operating modelApproval controls cover declared actions. A tool-server call outside a declared action is counted and capped per step, not gated. See how each is governed.
03 / The control center
Watch a real run, replayed from its record.
A crisis team of 31 agents worked day three of a national boycott at a fictional brewer. It listened, mapped the narratives, heard from every function and played three strategies forward. The CEO chose one 24 minutes in. About an hour later the response suite was ready, and the release waited overnight for a person.
What you see is the Control Center replaying that run from its record: 106 steps, $51.27 at list price, every agent, every decision and every file. The same work was forked at the CEO's card to play the other two strategies, and each path delivered its own pack. Nothing is staged, and any run you make can be replayed the same way. Each view answers a question someone accountable will ask.
Team
Who is on the team, what each agent may touch, and what each one did.
Floor
Where the work happened, desk by desk. Scrub to any moment of the run.
Score
Every step, who took it and what it cost, one lane per agent.
Board
Every file the team made, and the pack the lead handed over.
04 / How it works
Describe the organization. Alchemy runs it.
Four steps from a description to a delivered, verifiable result.
- 01
Describe
Write one manifest: the agents, what each may touch, which actions need approval and what counts as done.
- 02
Run
Hand it a piece of work. The lead plans and delegates, independent tasks run side by side, and you watch all of it from the control center.
- 03
Check
Before a run can finish, an agent that did none of the work opens every deliverable against your definition of done.
- 04
Record
A complete trail of every step: what each agent was given, what it did and what it changed. Encrypted on your disk, and anyone you share it with can check it has not been altered.
05 / What Alchemy does
Manage, observe, govern. In one place.
Agents need what any team needs: someone who sees the work, decides what matters and can prove what happened. Alchemy builds that in, so no team has to bolt it on.
Manage
Run a team of agents like a team.
Describe roles, reach and hand-offs once. Edit an agent with AI and keep every version, add skills with an administrator's approval, and start work from one box. Begin from 296 role templates or one of twelve ready projects.
Observe
See every agent at work, and every moment after.
The Control Center shows who is working, what is waiting on you and what each step cost. Replay any run, fork it from a decision, compare two runs, and let an AI audit read everything a project's agents did.
Govern
Nothing you mark happens without your people.
Mark pushes, filings, publishes and paid calls for a person. Give each agent only the files, tools and knowledge it needs. Cap spend per task, per project and per month, checked before every step.
Prove
The record exists before the auditor asks.
What every agent was given, did and changed, kept by the system rather than the agent. Encrypted on your disk, hash-chained, and exportable as a bundle anyone can verify without us.
06 / Where it works
Twelve ready projects. Eight industries. One way to govern them.
Each ships as a team you edit, not one you write from scratch. The work differs; the approvals, ceilings, independent checks and record are the same.
PR and communications
Crisis, launches and coverage
A crisis war room that works the whole response, a first-hour crisis desk, a release desk that checks every quote against the record, and a coverage report without inflated numbers. A person approves every release.
Marketing and strategy
Claims and memos you can stand behind
A green-claims sweep across markets and channels, and a red team for the market memo an investment committee will read. Every claim and figure traced to its source.
Finance, insurance and trade
Notes, renewals and refunds
An earnings note checked against the filings, a commercial renewal compared form by form, and an importer's tariff refunds traced to the last claimable dollar.
Software delivery and recruiting
Pull requests and submittals
A subcontractor's pull request checked before you pay for it, a vague ticket built into a tested change on GitHub and Jira, and a candidate submittal checked against the candidate's own CV.
07 / Trust
Control you can show, not describe.
Security
For the CISO
Alchemy runs on infrastructure you control, encrypted at rest under a key kept apart from the data. People sign in with named accounts or your single sign-on, with roles down to each project. Actions you mark wait for a person, and no agent can approve. Nothing is sent to Agiliti.
Finance
For the CFO
Set a dollar ceiling on every task, a ceiling on paid calls, and optional daily and monthly limits per project and for the whole account. Limits are enforced before each step and hold even when many runs start at once. Model use bills to your own Anthropic account, never marked up.
Platform
For IT
One self-hosted process on Windows, Linux or macOS, or in Docker, built on Anthropic's Claude Agent SDK. A command-line tool, a setup check, health and metrics endpoints and an operations runbook come with it. GitHub, Jira, Confluence, MCP tool servers and HTTP APIs connect as governed actions.
08 / Why we built it
A workforce needs an organization.
For two years the question in every large company was whether an AI agent could do a task. It can. Then we watched good pilots die in the meeting where finance asked what it would cost at scale and security asked who approved the actions. Nobody could answer, so nobody signed.
A tool needs an operator. A worker needs four things a tool never did: a role, a manager, a budget and a record. Without them a hundred agents is not a workforce. It is a hundred pilots.
You have seen it on this page: a team you can watch, a decision that waited for a person, a record kept by the system rather than the agent. We call it the fabric: the layer every agent action passes through, which asks who is acting, what it may touch and what it costs, and writes the answer down. Governance is the runtime itself, not a rule each team bolts on later.
We will always print the limit beside the claim. Claude is the only model. A tool-server call outside a declared action is counted and capped, not gated. We have no logos yet. When something breaks, you will hear it from us first, with the record attached.
If you have a pilot that worked and cannot get it signed, build the organization with us for eight weeks. We will run your first real piece of work together. You will read the record on Monday, and you can put your name on the result.