minarray
Menu

A team you run.
On the record.
Hard boundaries.

A small array of AI agents with real jobs: talk to each privately, bring them together, hand work between them. Every run is scoped, quota’d, and written to a record you can audit.

Request an inviteSee the boundaries

10 capabilities · 3 live · 7 modeled

  • lead_research
  • place_call
  • personal_assistant
  • web_search
  • receive_call
  • parse_document
  • generate_document
  • calendar_read
  • email_send
  • crm_write

minarray is a small array of agents you put to work : researching leads from a market brief and returning only what it can evidence, holding durable conversations that remember only what they were told to remember, and placing calls under human review. Every run is scoped to its capability, counted against a daily quota, and written to a durable record.

Ten capabilities. Three live. Nothing teased.

Live entries run in production today. Modeled entries are a roadmap stated as fact: specified, not built, no dates promised.

1.0 → how agents work
capability_registry10 entries
idscopestatus
lead_researchResearch a target market brief and return structured lead candidates, including company, contact, role, LinkedIn, and evidence, sourced from real web search results.LIVE
place_callPlace an outbound phone call to a lead via the telephony vendor.LIVE / GATED
personal_assistantHold a durable, voice-enabled conversation with the agent's owner: text or speech in, text or synthesised speech out, every reply a durable attributable run.LIVE / INTERNAL
web_searchSearch the public web for research material.MODELED
receive_callAnswer an inbound phone call routed to this agent.MODELED
parse_documentRead and structure an uploaded document or transcript.MODELED
generate_documentProduce a structured output document from context.MODELED
calendar_readRead the owner's calendar availability.MODELED
email_sendSend email on the owner's behalf.MODELED
crm_writeWrite lead/contact records back to the owner's CRM.MODELED

One run, on the record.

A real lead_research run from 2026-08-20, replayed from its record: the searches it ran, what it was not granted, and the verification that its results were real.

2.0 → the capability behind it
run_recordlead_research · 2026-08-20

  1. 01brief received

    “SaaS companies in Austin, TX, 10–50 employees, no dedicated ops function.”

  2. 02scope resolved

    • granted: lead_research, and nothing adjacent
    • not granted: place_call
    • not granted: personal_assistant

  3. 03executing

    web_search × 14: fourteen real searches, run by the research agent

    every claim carries its source URL · a field search didn’t surface stays null, never invented

  4. 04candidates returned

    • 01sources: YC profile · LinkedIn · press · confidence grade on record
    • 02sources: YC profile · LinkedIn · press · confidence grade on record
    • 03sources: YC profile · LinkedIn · press · confidence grade on record

    Two of the returned URLs independently re-fetched outside the run: both real, and the job postings the record cites are genuinely listed.

  5. 05record written

    2026-08-20 · durable · attributable · auditable

    counted: 1 run against the account's daily lead_research ceiling

    Company identities redacted for publication: they are real prospects, not customers. The run, its searches, and its verification are real.

The feature is what it won’t do.

Boundaries here are construction, not policy: a machine declining to act, on the record.

3.0 → boundaries
lead_research

It will not invent a fact.

Never invents, never upgrades. Every drop and downgrade is written into the run’s own warnings, so “it found less than it claimed” is visible rather than silent.

sanitize_result · verbatim from the validation code

drops:
a candidate with no company name, or no evidence
discards:
a LinkedIn URL that is not a linkedin.com address
downgrades:
to low confidence: no source URL backs this up
boundaries → enforcement
personal_assistant

It remembers only what it was told to.

Durable memory is admitted through one narrow, explicit tool call. It is never inferred silently from a transcript. Every stored fact traces to the turn that admitted it, and deleting one actually removes it from retrieval.

boundaries → enforcement
place_call

A person gates every engagement.

A person reviews every calling engagement before the first call is placed, and the work we can’t do lawfully gets declined.

boundaries → disclosures
  1. 4.1

    Invite-gated.

    There is no open sign-up: every account starts as a request a person reads.

  2. 4.2

    Human-approved.

    The people who build the agents review the requests.

  3. 4.3

    Quota’d.

    A real daily ceiling on every capability. Enforced, not courteous.

Request an invite

Small array. Real work. On the record.

Request an invitehello@minarray.com