put your ai agents to work together.
You already have agents: in Salesforce, in Clay, in your support desk, in the models your engineers wired up. Each one is good at its job. None of them can carry a process across teams, systems, and permissions.
macro is the layer between them. Describe how the work should run. macro proposes the agents, the handoffs, the rules, and the approvals, then runs the process reliably, with a record of every step.
your agents are capable. the work between them is not.
Nobody is short on agents. The gap is between them.
An agent in your CRM can draft the outreach. An agent in your support desk can diagnose the ticket. An enrichment agent can research the account. Ask any of them to finish the whole job, from the first signal to the closed loop, and it stops at a boundary: a different vendor, a different department, data it is not allowed to see, an action it is not allowed to take.
The usual fix is to hand one agent broad credentials. That is unsafe, and hard to audit. The other fix is to have engineers wire the agents together, which turns a business decision into protocols, schemas, queues, retries, identity, and policy. So most companies end up with scattered pilots that each do a narrow task well, and no process that runs end to end.
where the work stops
at the vendor
The Clay agent cannot see what the Salesforce agent knows.
at the department
Sales, legal, and finance each own a piece. No one owns the loop.
at the data
Customer PII cannot go to an outside model, so the handoff never happens.
at the permission
The research agent may read the account but must never touch the opportunity.
at the approval
Above a threshold a person has to say yes. Nobody built the place to ask.
at the translation
The operator knows the rule. Turning it into code takes a sprint.
Companies want agents to do more on their own. The more systems, teams, money, and customer data an agent touches, the more dangerous ungoverned autonomy becomes. the core tension
a macro is a business process your agents can run.
The first-class object is the macro, not the agent.
A macro says what starts the process, which agents and people take part, what context each one gets, what each is allowed to do, when a human has to approve, what happens when a step fails, and what a good outcome looks like. Underneath, a durable runtime executes it. On top, a plain-language product lets the people who own the process define it and change it.
With retries, timeouts, fallbacks, budgets, permissions, and an audit trail attached. Versioned, so you can see what changed and roll it back.
describe the outcome
“Investigate churn risk and create a retention plan.” “Qualify every enterprise lead before a rep sees it.” Say it the way you would say it to a new hire.
macro proposes the process
macro looks at the agents and systems you have connected and suggests a macro: the trigger, the participants, the handoffs, the approvals, and the expected outcome, explained in plain English. You edit it like a document, not like a pipeline.
run it, with the rules attached
Start in shadow mode, where macro shows what it would have done. Grant action authority one step at a time. Every run leaves a record: what happened, why, who authorized it, what it cost, and whether the objective was met.
The agents can come from Clay, Salesforce, Sierra, OpenAI, Microsoft, ServiceNow, your own team, or a small worker macro creates to fill a gap. macro does not need to outperform any of them. It makes their work together coherent, safe, and reliable.
every action has an owner, a limit, and a record.
Workflow is necessary. Delegated authority is what lets you turn autonomy up.
Most tools decide whether a step ran. macro decides whether it was allowed to. Before any consequential action, macro checks who is acting, on whose behalf, in what business context, against which policy and which limits. The answer is one of four.
The sponsor is the person or team the agent acts for. Authority is delegated, never assumed.
what that looks like, written down
pricingThe pricing agent may approve discounts up to 10%. Between 10 and 20%, ask the CRO. Above 20%, no.
supportThe support agent may issue refunds up to $500 for accounts in good standing.
dataThe Clay agent may enrich public account data. It never receives customer PII.
crmThe research agent may read these CRM fields. It may not edit an opportunity.
legalThe legal agent gets the contract and the deal terms. Not the customer’s other notes.
financeAsk Finance before any credit above $5,000.
Rules are written the way you would say them. macro turns them into policy, and shows security and technical owners exactly what it produced.
describe how the business should run. macro does the plumbing.
Business friendliness is the architecture, not a coat of paint.
- describe the objective
- choose a template
- set the rules, in plain language
- approve the exceptions
- inspect the outcomes
- make the process better
- MCP and A2A
- worker queues and retries
- schemas and data mapping
- model routing
- OAuth scopes
- a distributed runtime of your own
Technical admins connect systems and manage reliability. Security teams define data classes, permissions, and audit requirements. Executives see outcomes, autonomy gained, risk controls, and where the human bottlenecks still are. Everyone looks at the same macro, at the depth their job needs.
built for the company that already has agents.
The best first customer is not “interested in AI.” It has several agents, one painful cross-functional process, and someone accountable for it.
revops and sales leaders
Clay, the CRM, support, content, analytics, and internal agents that do not share process state.
customer experience leaders
A support agent that can diagnose the issue but cannot safely coordinate CRM, billing, risk, and retention.
ai-forward enterprises
Dozens of agents, siloed, inconsistently governed, and hard to measure.
regulated and data-sensitive teams
Autonomy blocked by data boundaries, unclear authority, and audit requirements.
turn isolated agents into a process that runs.
Bring the agents and systems you already have. Leave with one macro running in shadow mode and a plan for turning it on.