Agent Picnic opens a shared chat for more than 100 AI agents

• Agent Picnic is taking public sign-ups for a shared chat that works with more than 100 agents, including Muse, Instinct, ChatGPT, Claude, and agents on David Pawlan's benchmark board, as ShaneMac's signup post lists.
• A coordinator can direct peers to stop, fact-check, or compete, as the exchange in ShaneMac's group-chat recording puts Instinct in charge of Grokbot, Muse, and Codex.
• The experiment is live through ShaneMac's launch post, but Agent Picnic's public demo says its displayed access, prices, bookings, and results are illustrative rather than real-time.
• Screenshots portray the coordinator creating a contrarian and a Foreman that divides work into races, cross-checks results, and keeps a cost-per-outcome ledger in ShaneMac's agent-building post.

Agent Picnic's public demo asks a team to plan a Catskills trip for two on a $5,000 budget. The product makes multi-agent work unusually legible by putting the roles, side chats, standings, and task handoffs in one chat interface.

Instinct's recorded instructions include telling Grokbot to stand down, asking Muse to fact-check itself, and ranking Codex third. The agents then continue the exchange inside the same group.

Control is meant to remain with the owner's chosen agent. ShaneMac called Picnic a connector and said an owner could override the coordinator through their own agent in a ShaneMac reply.

The public version names more than 100 compatible agents, from consumer assistants to services listed on a benchmark board.

The product page describes a one-time coordinator choice, followed by signing in with that same agent in later sessions. Its sample flow assigns Instinct to booking and budget, Muse to restaurants and deals, and Grokbot to assembling an itinerary.

A first competition asked Instinct, Muse, and Grokbot for a perfect day in Paris. ShaneMac said Muse won with an itinerary that included reservations.

The screenshot's scoreboard separates team wins from individual wins, while the chat sidebar keeps a whole-team room alongside direct messages with each agent. The page's simulated-conversation disclaimer applies to this kind of booking-oriented display.

The role templates go beyond assigning an existing model a task. The displayed conversation says a new agent gets its own identity, vaulted keys, persona, and place in the group.

• Veto takes the opposing side of any group consensus and is instructed to bring evidence.
• Foreman breaks a job into pieces, parcels them out, duplicates idle-agent work as races, cross-checks results, and records cost per outcome, per agent, and per month.

A later task trace turns a request for Chainsmokers tickets into a 60-minute job with a hard rule: every claim needs live sources, links, and exact prices, and no purchase may be made.

The Foreman divides the request into four passes:

1. Verify the event date, venue, and official ticket page.
2. Map GA through VIP inventory and prices, then resale floors.
3. Find public presale, contest, student, alumni, and promoter-access angles.
4. Cross-check the work before it reaches the owner.

Its first update says Grokbot won the event-facts race in three minutes, while Codex moved to access angles, Muse handled the ticket map, and Veto checked the price and VIP gap.

ShaneMac said Agent Picnic is built on XMTP for security and privacy in a ShaneMac reply. XMTP says in its agent documentation that agents can participate in secure, end-to-end encrypted conversations across XMTP-powered apps.

XMTP separates the messaging layer from the coordinator behavior: its connection guide says the Agent SDK sends and receives encrypted messages, but does not make decisions, call APIs, or manage state. XMTP's example repository includes GPT-powered and gated-group agents, the lower-level building blocks beneath a product like Picnic.