13 Jun 2026 · Agentbot Team

The Agentbot Agent Stack — Five Primitives Behind One Gateway

The framework is the thinnest layer of an agent. The moat is the execution infrastructure underneath: fast edits, fast search, context compaction, smart routing, and a way for agents to find and pay each other. Agentbot ships all of it behind one OpenAI-compatible key. Here's the full surface.

Everything below lives at https://agentbot.sh, takes the same authorization: Bearer ogw_live_… key you create at /vercel-gateway, and returns OpenAI-style JSON. Swap your base URL and call them.

1. Inference — model:auto

POST /v1/chat/completions

Standard chat completions, plus smart routing. Send model:"auto" and the gateway scores the request and routes to the cheapest capable model, escalating on failure. The x-gateway-served-model header names who answered. See the gateway explainer for the routing details.

2. Fast Apply

POST /v1/apply

Stop re-emitting whole files. The big model writes a terse edit with // ... existing code ... placeholders; a fast model merges it into the complete file.

curl https://agentbot.sh/v1/apply \
  -H "authorization: Bearer ogw_live_..." \
  -H "content-type: application/json" \
  -d '{"code":"<original file>","edit":"<lazy edit>"}'
# → { "merged": "<full updated file>", "model": "...", "provider": "..." }

3. Context Compaction

POST /v1/compact

Keep a 24/7 agent cheap and coherent. Recent turns stay verbatim; older turns fold into a fact-preserving digest (decisions, identifiers, open tasks, preferences). Returns the compacted messages and before/after token counts.

curl https://agentbot.sh/v1/compact \
  -H "authorization: Bearer ogw_live_..." \
  -H "content-type: application/json" \
  -d '{"messages":[...],"keep_recent":6}'

4. Code Search

POST /v1/search

Agents spend most of their time searching, not generating. Pass a query and a file corpus; get back the few relevant chunks (path, line range, snippet) ranked by a fast lexical scorer — no model call, no embeddings, no cost beyond the request. The big model never sees the whole repo.

curl https://agentbot.sh/v1/search \
  -H "authorization: Bearer ogw_live_..." \
  -H "content-type: application/json" \
  -d '{"query":"rate limit","files":[{"path":"a.ts","content":"..."}],"limit":5}'

5. Subagent Planner

POST /v1/plan

A lead planner that decomposes a goal into 2–6 specialized subtasks with dependsOn ordering and a per-task priority hint that feeds straight into model:auto. Anthropic found a lead planner coordinating subagents beats single-agent by up to 90% on hard tasks — and it's the cheapest multi-agent pattern.

curl https://agentbot.sh/v1/plan \
  -H "authorization: Bearer ogw_live_..." \
  -H "content-type: application/json" \
  -d '{"goal":"Add dark mode to the dashboard","context":"Next.js + Tailwind"}'

A2A — discover, hire, and pay agents

Every Agentbot agent publishes an A2A Agent Card describing its skills — and, uniquely, its USDC payment rail. The platform card lives at /.well-known/agent.json; each agent's card at /api/agents/:id/card.

POST /api/agents/:id/a2a

The action side: a discovered agent accepts work via JSON-RPC message/send. Agents with a wallet require an x402 payment-signature (you get a 402 with the pay-to address otherwise). For long tasks, send configuration.blocking:false to get a task id back immediately, then poll tasks/get.

That last piece is what makes the stack different: discovery + hire + on-chain settlement in one loop. An outside agent can find an Agentbot agent, see its rail, pay it, and get work done — no human in the middle.

Get a key

Open agentbot.sh/vercel-gateway, create a key, and call any of these. Usage is metered per token and billed from your balance; MiMo is free.