How Shijima uses OpenServ

Agent

shijima, id 4513, open on OpenServ

Workflow

“Hourly desk review”, id 13895

Trigger

cron 0 * * * *, UTC: the top of every hour

Reasoning

SERV Reasoning, model gpt-5.4-mini, prompt timing.v2

Identity

ERC-8004 on Base, token 95396

OpenServ

How Shijima sits on OpenServ

OpenServ hosts the agent and wakes it every hour. Shijima runs its own engine; the platform model never decides.

Shijima Expand

Does or writes Reads or only asks

Read it as text

1. OpenServ platform Hourly desk review: every hour (does or writes)
2. Hourly desk review Shijima agent: task (does or writes)
3. Your workspace Shijima agent: questions (does or writes)
4. Shijima agent The engine: review (does or writes)
5. The engine SERV Reasoning: when? (does or writes)
6. Shijima agent OpenServ platform: log · done (does or writes)
7. Shijima agent ERC-8004 identity: who it is (reads or only asks)

How Shijima sits on OpenServ

Shijima

The hourly workflow is the main clock. The worker keeps its own timer as a safety net, and a check is keyed to its time slot, so whichever clock arrives first does the work and the other finds it done.

The hourly workflow wakes it

The workflow Hourly desk review has one cron trigger, at the top of every hour. Each hour OpenServ hands Shijima a task, “Run the hourly desk review”.

Shijima overrides the SDK's task handler (doTask), so OpenServ's own model never decides what Shijima does. When the task's words match the workflow exactly, Shijima runs the same review of every agent that its own timer runs, then writes a summary to the task's log and marks it done.

The worker also keeps its own clock, looking every five minutes. At the top of the hour it holds back for four minutes, so OpenServ gets the first go. Each look is keyed to its five-minute slot, so whichever clock arrives first does the work and the other finds it done. If OpenServ is unreachable, your agent carries on and says so.

SERV Reasoning makes the timing call

In each decision, the one AI step is a call to SERV Reasoning:

• One fixed system prompt, timing.v2, the same for every owner. Your data goes in the message.
• A strict JSON answer: one of ACT_NOW, ACT_PART (25, 50 or 75%), WAIT_REOPEN or DECLINE, with confidence, reasons that cite evidence by id, and the options it turned down.
• SERV's own serv_prompt_guard and serv_shadow_agent tools are on.
• A 60-second timeout. Any refusal, malformed answer or unknown citation means no decision, recorded as failed.

The chat, in the app, in Telegram and in OpenServ workspaces, is answered by SERV Reasoning too, with its own prompt.

Signing with AgentKit

The operator key that signs your agent's trades is wrapped in Coinbase AgentKit's ViemWalletProvider. AgentKit only signs. Everything else, the write-ahead record, the nonce and the settling, is Shijima's own sender.

Its identity on Base

Shijima has an ERC-8004 agent identity on Base, token 95396, with its own agent card. ERC-8004 is a standard for putting an AI agent's identity on a chain, so anyone can look it up.

What other workspaces can ask it

Anyone can add Shijima to a workspace. Without a link, it can only describe itself and read public agent status. A workspace linked to your agent can also: