Researcher
Investigates, checks sources, and identifies uncertainty.
INSIDE OUR OWN AI SYSTEM
A team of agents.
A system of responsibility.
See how a brief becomes useful work, how decisions are remembered, and where a human stays in control.
THE DESIGN CHALLENGE
Different work.
Different responsibilities.
Research, code, and communication need different tools and permissions. Our founder’s AI assistant gives each job a specialist, shared context, and clear boundaries.
01 / THE HANDOFF
One brief. A clear path.
- 01
You set the goal.
Define the job and what a useful result looks like.
HUMAN - 02
Work gets routed.
The orchestrator sends a scoped task to a specialist.
ORCHESTRATOR - 03
A specialist delivers.
Results land in its own topic. The board tracks progress.
AGENT - 04
You make the call.
Approve calls and publishing. Decide what code gets merged.
HUMAN
02 / THE SPECIALISTS
A role. A useful result.
Sourced brief ↗
Developer
Implements changes and runs checks in real repositories.
Caller
Makes approved outbound calls, discloses it is AI, and reports the outcome.
Growth lead
Turns a project brief into content and campaign drafts.
03 / MEMORY
Keep the decisions.
Keep the context.
Conversation↓Memory curator
Proposes useful facts↓Human review ✓
Proposes useful facts↓Human review ✓
Durable memory has private and shared scopes. A proposed memory is saved only after review.
PostgreSQL-backed context04 / AUTHORITY
Every action
has a boundary.
Policy checks the agent, tool, and action. Access follows the role.
✓
Allow
Proceed within granted permissions.
↗
Ask
Bring the action to a human.
×
Deny
Stop an action outside the boundary.
THE PRINCIPLE WE BRING TO YOUR SYSTEM
Start with a job.
Design for responsibility.
Define the tools, memory, result, and approvals. Add more agents when the work needs separate roles.
Explore custom AI systemsFROM THE IMPLEMENTATION
Decisions you can inspect.
Research before the call.
When a call needs current facts, the caller requests research through the orchestrator. The returned facts go into the call’s talking points before approval.
ChosenUse the existing researcher.AlternativeGive the caller its own research tools.TradeoffMore handoffs and latency; clearer separation of roles.
Condensed from our design note and current caller role definition.Recognize a repeated report.
End-of-call reports use a key based on the call and event type. A stored duplicate key prevents another handler invocation. Requests with an invalid webhook secret are rejected.
New event key→Handle
Stored event key→Skip
Invalid secret→Reject
Architecture notes & provenance
This is VERV’s founder’s own multi-agent assistant, built with TypeScript services, a Telegram interface, and PostgreSQL-backed memory. The architecture and role definitions were reviewed in September 2026. Every outbound call requires approval; the caller introduces itself as AI and cannot agree to payments or contracts. Examples illustrate responsibilities without reproducing private conversations.