AgentCompile is for agents that do real work through tools: look something up, then change something. Wherever an agent does the same jobs again and again, those jobs can run compiled, and your agent's model is kept for everything else.
Lookups, then a write. The industry changes; the shape doesn't.
Where it works today
- Customer support and operations agents: refunds, changes, cancellations and account updates, for airlines, retailers and telecoms.
- Workflow agents across SaaS tools: email, calendar, CRM and ticketing.
- Voice agents: the same jobs, over the phone.
Close to what we do today
- Data and SQL agents: the same questions every week, with new dates and regions.
- Document and extraction agents: invoices, receipts and contracts into fields.
- Internal ops and IT agents: access resets, new accounts, key rotations.
Where we're heading
- Computer-use and desktop agents: the same jobs, clicked through native apps.
- Multi-agent handoffs: one agent choosing which specialist takes over.
What we leave alone, on purpose
Some agent work has almost nothing that repeats, and AgentCompile doesn't try to compile it. That's a choice, not a gap.
- Open-ended research and writing. Every run is different: a new question, new sources, a new answer. There is no job that comes back, so there is nothing to compile beyond the tool calls along the way.
- Creative or one-off reasoning. A strategy memo, a new design, a hard question asked once. The only run is the one in front of it, and that thinking is exactly what your agent's model is for.
For work like this, AgentCompile stays out of the way: every call goes to your agent, unchanged. Thinking is for the new, and this is all new.
The test: does the work come back?
Industry is the wrong way to sort agents. A support agent at a bank and one at a retailer have more in common with each other than with a research assistant at either company. What matters is the shape of the work: does the same job come back, again and again, with new details each time?
- Same job, new details: a refund for a different order, an invite for a different meeting, a report for a different week. This is where AgentCompile works.
- Tools, not just text: the agent looks something up, then changes something. The lookups and the write are what repeat.
- Volume: a job done ten times a day is a convenience to compile. A job done ten thousand times a day changes how your agent runs.
If the job comes back, it can be compiled. If it never comes back, it shouldn't be.
How to read the rest of this blog
Each kind of agent below has its own post: what its repeated jobs look like, what changes when they run compiled, and what stays with your agent. Support, workflow and voice agents are where AgentCompile works today. Data, document and IT agents are close to it. Desktop agents and multi-agent handoffs are where we're heading, and we say so plainly in those posts.
Questions we get
- What if my agent does a bit of everything?
- Most do. AgentCompile only takes the jobs that repeat and leaves the rest alone, so a mixed agent still gets the benefit where the repeats are, and nothing changes anywhere else.
- How do I know which jobs my agent repeats?
- That's the start of every pilot: show us your agent's logs, and we show you which jobs repeat and how often.
- Is there a minimum volume?
- There is no line we draw. The more often a job comes back, the more it matters that it runs compiled.
Each kind of agent has its own post. Read the blog