Primitives
The Amdahl learning model processes your data, analyzes it, models what wins with your buyers and optimizes the output of your agents against that model. Two calls reach it: the optimize call writes and evaluates outreach, and Search returns rows and quotes from your processed customer data. Research answers one question on top of them.
Amdahl is an applied machine learning company for GTM agents. Its learning model does four things with your data, and two API calls, the optimize call and Search, are how your agent uses it. Everything else on the API is built from those two or manages the data behind them.
The learning model
- It processes and structures your data. Connections sync your CRM, call recorder and email into one searchable shape, speaker-attributed and joined to the deal. Classifiers tag every message across 10 dimensions before any query runs. You never send this data with a request: it is already on the server, scoped to your workspace and to what your credential may read.
- It analyzes it. Search counts, filters and groups that data, and finds what customers said in their own words.
- It models what wins with your buyers. An evaluation measures a message against a fixed set of criteria and, with
evidence: "workspace", against your customers' own conversations. Every claim nothing in their words backs is listed. - It optimizes your agent's output against that model. The optimize call tries versions of a message, evaluates each, and returns the best one in the sender's voice.
optimize_promptdoes the same for the prompt your agent writes from.
Your agent writes. Amdahl optimizes it to win, and it works with any agent: over REST, over MCP, or from the amdahl CLI.
The optimize call
POST /messages/optimize (the optimize_message tool over MCP) takes what you know about the recipient, in context, and an optional draft. It returns one email or LinkedIn message in the sender's voice on the same call, usually in 1 to 2 minutes. When you send a draft and nothing it tries beats it, it returns the draft unchanged.
- It needs no connected data. Without your data, it evaluates how a message reads and how clearly it asks. That is why outreach is the first result a new workspace can get.
- It says only what it can stand behind. A claim must come from a fact you sent or from evidence Amdahl read. A version that adds a figure or a claim nothing states is discarded.
- It reads your data when you ask. With
"evidence": "workspace"on a workspace with synced conversations, it also evaluates the message against what your customers said, and lists the claims nothing in their words backs. - It sends nothing. It hands back text, and nothing in your tools or your data changes. Where the text goes is your decision.
- It also evaluates without writing. With
mode: "grade"it evaluates a draft on five criteria and writes nothing: a pass or fail per criterion with a reason, and a 1 to 5 score withpassedandtotalbeside it.optimize_promptevaluates a reusable prompt by the emails it writes, and inimprovemode returns a revision only when it wrote better email on held-out recipients. See What an evaluation measures.
Scope: messages:execute. Start with Optimize a cold email; the references are Optimize a message, Evaluate a message and Improve a prompt.
Search
| Search | |
|---|---|
| Call | POST /search/query |
| Shape | Synchronous, one blocking call (a 15 second ceiling on the default path) |
| You send | Typed filters, a plain-language question, a meaning to match, or a quoted phrase |
| You get back | Rows or utterances, the SQL or terms that produced them, and a corpus block naming the store that answered |
| Scope | data:read |
| Changes your data | Never |
| Reference | Search |
Search answers a question about your conversations. A call on an empty workspace has nothing to read, so check GET /search/overview before you trust an empty result. The reference explains the four search modes and which one answered (mode_ran); read that field before you compare two results.
Choosing between them
| You want | Reach for |
|---|---|
| One outbound message from what you know, now, with nothing connected | The optimize call |
| That, plus every claim in the message evaluated against what your customers said | The optimize call with evidence: "workspace" |
| A number, a list, a slice of deals or conversations | Search, mode: "filter" when you know the fields; plain language otherwise |
| What customers said about a topic, in their words | Search, mode: "semantic" with hydrate: true |
| Whether a specific word or name was said | Search with the term in double quotes |
| Every criterion of a draft evaluated, with a reason for each | The optimize call with mode: "grade" and evidence: "workspace" |
| A send-or-hold decision on copy before it ships | The same evaluation, read on grade.passes_bar |
| A better prompt for next time | optimize_prompt, in its default improve mode |
| One question that needs several reads and a written synthesis | Research, which runs the reads for you |
Research
This is a convenience built on the two calls. It reads nothing Search and the optimize call cannot.
- Research runs a read-only agent server-side that reads your processed customer data and returns one cited answer. You get a handle back and poll. It reads only your own data and does not search the web.
Use it when you would otherwise write the loop yourself. Skip it when you want control flow in your own code: two Search calls and an evaluation are cheaper and easier to test than a research job that does the same thing.
Data behind the calls
Connections manage the CRM, call recorder, email and other sources the learning model reads: what can be connected, what is connected, and how healthy each sync is. They change what Search and the optimize call can read; they do not answer questions.
Two ways to call them
The same operations are reachable over REST (https://app.amdahl.ai/api/platform/v1, with an API key) and over MCP (https://app.amdahl.ai/mcp, see Connect your agent). Over MCP, each operation is its own tool: Search is search_customers; writing and evaluating an outbound message is optimize_message; evaluating or improving a prompt is optimize_prompt and get_prompt_run; recording what happened to a message is record_message_feedback; a research question is search_customers with route: "research", then get_research; and connections are list_connections, connect_source and their siblings. A get_ or list_ tool only reads and changes nothing. A tool that starts a job or makes a change may ask the user to approve the call.
Next
- Optimize a cold email: outreach on one recipient, read end to end.
- Evaluate a cold email: one evaluation read end to end, then the prompt behind it.
- Evidence and trust: the fields that say how far a result can be trusted.
- Patterns: the call sequences that work.
- Use cases: GTM jobs mapped to these calls.