My First Agent

{{ projectName }} ›

🪙 {{ runCreditsLabel }}

Low on credits — see plans →

WORKSPACE VIEWS

Agent Session

Memory & Dreaming

Family Dashboard

📜 My Sessions

Outcome & Graders

BUILD & INSPECT

Agent YAML

Activity Logs

ADMIN

Site Admin

Backend…

ACCOUNT

Sign out

This backend is running in dev mode — unsupported. Set AUTH_PASSCODE on the server and reload.

Cant reach the server — youre offline and your changes are saved locally.

when Run is clicked

set goal to

workspace:

repeat until answer is ready or {{ triesHex }}

think about the next step

TOOLBELT — TAP TO ADD OR REMOVE · THE AGENT PICKS EACH LOOP

+ add tool

look at what came back

human in the loop: the agent checks with me

go around again ↻

done looks like {{ outcomeBadge }}

show the final answer

The agent runs this loop over and over until the goal is done. Tap Code to see the real code →

Monaco editor · generated live from your blocks

// my agent = a loop with two ways to stop
import { query } from "@anthropic-ai/claude-agent-sdk";

const goal = "find the 5 weirdest deep-sea creatures and write a fact file";
const tools = ["WebSearch", "Read", "Write"];  // allowed, not ordered!
const helpers = { "fact-checker": { description: "Double-checks every fact" } };
const workspace = { files, skills, memory };  // the agent's desk

for await (const msg of query({  // ↻ until answer is ready…
  prompt: goal,
  options: {
    allowedTools: tools, agents: helpers,
    maxTurns: 10,            // …or max turns reached
    canUseTool: askMeFirst   // human in the loop
  }
})) workspace.memory.push(msg);  // save each step into memory

Match the colours: every block on the Blocks tab is one piece of this real code.

THE REAL REQUEST — WHAT RUN ACTUALLY SENDS

// POST /v1/agents  (then POST /v1/environments, POST /v1/sessions, user.message)
{
  "name": "Deep-Sea Detective",
  "model": "claude-sonnet-5",
  "system": "You are a friendly kid-safe agent. Mission: find the 5 weirdest deep-sea creatures and write a fact file",
  "tools": [{
    "type": "agent_toolset_20260401",
    "default_config": { "enabled": false },
    "configs": [
      { "name": "web_search", "enabled": true,
        "permission_policy": { "type": "always_ask" } },
      { "name": "read", "enabled": true, "permission_policy": { "type": "always_ask" } },
      { "name": "write", "enabled": true, "permission_policy": { "type": "always_ask" } }
    ]
  }],
  "skills": []
}
// then the outcome block kicks it off (no user.message needed):
{ "type": "user.define_outcome", "description": goal,
  "rubric": { "type": "text", "content": "done looks like…" }, "max_iterations": 3 }

The exact JSON your blocks compile to — goal → system, toolbelt → configs, human in the loop → a check_with_me custom tool the agent must answer through you.

AGENT SESSION

👁 Shared session — view only

This session isn’t available — it may be private, removed, or the share link may have expired.

Press Run ▶ or send a message to start the mission…

{{ row.text }}

AGENT

{{ runPreview }}▌

permission: let the agent use {{ runPendingName }}?

the agent asks you: {{ runPendingQuestion }}

All done — {{ runStopText }}

A parent can unlock more with a bigger plan — See plans →

🎉 {{ firstSuccessText }}

AGENT CONFIG — SYNCED WITH YAML

GOAL — WHAT THE AGENT SHOULD DO

Synced live with the coral block — this is the mission Run sends.

NAME

DESCRIPTION

MODEL

SYSTEM PROMPT

Loaded from the YAML page when saved — edits here are what Run sends.

TOOLBELT CONFIG

human in the loop — the agent pauses and checks with you (its own block, not tied to tools)

MCP SERVERS

default always_ask · try deepwiki — free, no login

{{ mc.url }}

Add a tool server

Pick a server your agent can use. The Playground always works — perfect for your first try.

{{ e.tools }}

HUMAN IN THE LOOP

Gives the agent a check_with_me custom tool. When it calls the tool, the session pauses (requires_action) until you answer — your reply goes back as user.custom_tool_result and becomes part of the agent's context. Not tied to web_search / read / write permissions.

enabled — show the block and give the agent the tool

WHEN SHOULD IT CHECK WITH YOU? — becomes the tool's description

Heads-up from the field: humans approve ~93% of prompts and tune out fast — steer the agent to ask FEWER, higher-stakes questions, not more.

WORKSPACE

FILES

synced to the sandbox via the Files API when live

{{ f.path }}

👁

↓

{{ fileViewPath }}

SKILLS

stored locally, attached at agent create

{{ sk.name }}

MEMORY

syncs to your local file system

{{ m.path }}

DREAM REPORT — dream-log.md

🧠 MEMORY & DREAMING

DREAM EFFECTIVENESS — is dreaming curating or bloating?

files

size

read the dream report →

MEMORY — what your agent has learned {{ memCounts }}

{{ grp.folder }} · {{ grp.count }}

👪 FAMILY — PARENT DASHBOARD

CHILDREN — profiles, controls & spend this month

models

+ add a child profile

RECENT RUNS — what they cost

No runs yet this period — costs appear here after each run.

📜 MY SESSIONS

YOUR RUNS — NEWEST FIRST · TAP A ROW TO OPEN ITS REPLAY

shared

No sessions yet — press Run ▶ and your runs will show up here.

🛠 SITE ADMIN — SITE OVERVIEW

computed $ split — {{ adminCostSplit }}

Dollar figures are computed from usage events, not Anthropics invoice — reconcile against the org-level Usage/Cost API.

USERS — COMPUTED SPEND PER ACCOUNT

BY MODEL

PLANS

{{ p.name }} · {{ p.line }}

RECENT RUNS (ALL ACCOUNTS)

ALL SESSIONS (EVERY ACCOUNT) — TAP TO OPEN

{{ asr.model }} · {{ asr.status }} · {{ asr.meta }}

shared

No sessions recorded yet.

GLOBAL TEMPLATES

No global templates yet — create the first one.

Add MCP from Library:

OUTCOME — WHAT "DONE" LOOKS LIKE

Sent as user.define_outcome — a separate-context grader scores each rubric line and the agent iterates until satisfied or out of tries.

RUBRIC — MARKDOWN CRITERIA, ONE GRADEABLE LINE EACH

RUBRIC SOURCE

TRIES

default 3 · max 20

GRADER LIBRARY

checked graders AND together into the rubric · tap i for details

emits → {{ g.emit }}

COMPOSED RUBRIC — SENT WITH user.define_outcome

{{ composedRubric }}