Try for free
Workers

Agent runtime

The satellite that hosts the 'agent builder' surface. One Durable Object per (agent id, conversation id). Each instance drives its own inference loop, dispatches tool calls (built-in, unified-connector, custom MCP), and persists state across turns.

Updated today

Overview

At a glance
Worker
ultron-agent-runtime
Durable Object
AgentInstance
DO key
(agent_id, conversation_id)
D1
ultron-agent-builder-db (agent defs, integrations, deployments)
R2
Artifacts (uploaded files for RAG, deployment bundles)
Cron
* * * * * (every minute)
Model traffic
Routed through an AI Gateway slug for cache, logs, rate limits
Vectorize indices
Per-agent agent-rag-<agent_id>, created on demand

Where the chat-v2 route in the app is the user-facing inference loop, the agent runtime is the satellite that the "build your own agent" feature delegates to. The two run on different schemas, with different integration projects, so a misbehaving experimental agent can never touch the production app state.

Worker shape

wrangler.toml (key bindings)toml
1name = "ultron-agent-runtime"
2main = "src/index.ts"
3
4[[durable_objects.bindings]]
5class_name = "AgentInstance"
6name = "AGENT_INSTANCE"
7
8[[d1_databases]]
9binding = "AGENT_DB"
10database_name = "ultron-agent-builder-db"
11
12[[r2_buckets]]
13binding = "AGENT_ASSETS"
14bucket_name = "ultron-agent-builder-assets"
15
16[ai]
17binding = "AI"
18
19[vars]
20AI_GATEWAY_SLUG = "ultron-agent-builder"
21COMPOSIO_PROJECT_NAME = "ultron_agent_builder"
22COMPOSIO_API_BASE = "https://backend.composio.dev/api/v3"
23
24[triggers]
25crons = ["* * * * *"]

AgentInstance DO

One DO instance per (agent, conversation) pair.

src/index.ts (DO sketch)ts
1export class AgentInstance {
2 state: DurableObjectState
3 env: Env
4
5 async fetch(req: Request) {
6 // Routes:
7 // POST /turn start a new assistant turn
8 // POST /tool-result return a tool call result and resume the loop
9 // GET /transcript read the full transcript
10 // POST /reset drop conversation state
11 }
12
13 // Driven by /turn:
14 async runLoop(input: string) {
15 let messages = await this.loadMessages()
16 messages.push({ role: 'user', content: input })
17
18 while (true) {
19 const resp = await this.callModel(messages)
20 if (resp.tool_uses?.length) {
21 const results = await this.dispatchTools(resp.tool_uses)
22 messages = messages.concat(resp, results)
23 continue
24 }
25 await this.persist(messages.concat(resp))
26 return resp
27 }
28 }
29}
Note
State lives in the DO's own SQLite storage. Each agent's conversation is a separate DO instance, so writes are serial per agent and reads are local.

Tool surfaces

What an agent built in the runtime can reach for.

SurfaceWhat it gives
Built-in primitivesBrain read, web search, calculator, document generation. Same shape as chat v2's native tools but scoped to the agent's policy.
Unified connector (per-agent)OAuth surfaces under the siloed agent-builder integration project. Gmail, Slack, Notion, Google Workspace, etc. Each agent has its own connected accounts.
Custom MCP serversAgent owner can register an MCP endpoint with credentials; the runtime connects, lists tools, exposes them to the model.
Workspace skillsSkills registered to the agent's owning workspace appear as callable tools in the loop.

Per-agent scheduling

How an agent fires itself on a schedule.

Each agent definition row in AGENT_DB can carry a cron expression plus a cron_next_run_at timestamp. The worker's 1-minute scheduled handler scans for agents whose cron_next_run_at <= now() and enqueues a turn for each, then advances cron_next_run_at.

src/index.ts (scheduled handler concept)ts
1async scheduled(_event, env) {
2 const due = await env.AGENT_DB.prepare(
3 `SELECT id, owner_user_id, cron_expr
4 FROM agents
5 WHERE cron_expr IS NOT NULL
6 AND cron_next_run_at <= datetime('now')`
7 ).all()
8 for (const row of due.results) {
9 const stub = env.AGENT_INSTANCE.idFromName(`${row.id}:cron`)
10 const obj = env.AGENT_INSTANCE.get(stub)
11 await obj.fetch(new Request('https://internal/turn', {
12 method: 'POST',
13 body: JSON.stringify({ source: 'cron' }),
14 }))
15 // advance cron_next_run_at by the cron interval
16 }
17}

Per-agent RAG

Vector indices are created per agent on demand.

When the agent owner toggles RAG on, the runtime calls the platform's vector-index management API to create a new index named agent-rag-<agent_id>. Uploaded files in AGENT_ASSETS are chunked + embedded and inserted into that index. Inference-time, the inference loop queries the index for relevant chunks before each model call.

File map

cf-workers/ultron-agent-runtime
wrangler.toml
src
index.tsDO class, fetch + scheduled handlers, AI Gateway client
tools.tstool dispatch + connector adapter
rag.tsembedding + index management
migrations
0001_agents.sql
0002_integrations.sql