sapiom-agent-builder（sapiom Agent 构建器）

Sapiom Agent Builder

Build and deploy autonomous agents powered by @sapiom/fetch SDK and Sapiom service gateways.

When to Use

• User says "build me an agent that..."
• User wants to "automate X on a schedule"
• User wants a "bot that checks/monitors/scrapes/summarizes"
• User wants to "alert me when X changes"
• User wants to deploy code that runs autonomously

Already have a working script and just want to deploy it? Use the sapiom-deploy skill instead.

Process

Phase 1: Parse Intent

Determine from the user's request:

1. What should the agent do? — Monitor a website? Summarize content? Research a topic?
2. How often? — Every 5 min, hourly, daily, weekly? On-demand?
3. What output? — Log to console? Send a webhook? Store results?

Pick a compute pattern:

For agents that iterate over a list of items, consider the orchestrator + workers pattern: a lightweight scheduled job handles scraping and diffing, then triggers a separately deployed batch job with new items as task payloads. This processes items in parallel instead of sequentially.

Full deploy API details: references/deploy.md

Map the intent to a service pattern:

Phase 2: Select Capabilities

Read the reference for each service the agent needs. Only read what you need. For each service, ask: does the agent actually need this, or is it just nice-to-have? Extra service calls add cost, latency, and failure points — and if a service returns low-quality data, downstream LLM analysis will be wrong too.

Phase 3: Get API Key

The user needs a Sapiom API key. Two options:

1. MCP tool (if sapiom_create_transaction_api_key is available): Tool: sapiom_create_transaction_api_key Parameters: {name: "{agent-name}-key", description: "API key for {agent-name}"} The plain key is shown only once — save it immediately.
2. Manual (if no MCP tool): Direct user to https://app.sapiom.ai/settings to create an API key.

If the user already has a key, skip this step.

Phase 4: Write the Script

Start from the appropriate template:

Read the template, then adapt it:

1. Copy the SDK setup pattern (same for all agents)
2. Add gateway calls from the reference docs
3. Wire the steps together: gather → process → act
4. Add error handling:

- Every safeFetch call should check res.ok and include the response body in the error: throw new Error(\... HTTP ${res.status}: ${await res.text()}) - If the agent iterates over a list of items, wrap per-item work in try/catch so one failure doesn't abort the entire batch. Log the error and continue.

Script requirements for deployment:

• Entry point: index.js
• Must have package.json with @sapiom/fetch in dependencies
• Must have scripts.start: "node index.js"
• Agent reads SAPIOM_API_KEY from process.env
• Keep it simple — one file, no build step
• Batch job workers use const {blStartJob} = require("@blaxel/core") and call blStartJob(myFunction) where myFunction(args) receives each task's payload. Add @blaxel/core to dependencies. Scheduled jobs and standalone scripts use a normal main() function.

If the agent uses Postgres, always include an agent_runs table to track execution history. See the "Run Tracking" pattern in references/database.md. Every agent run should:

1. INSERT a row into agent_runs with status running at the start
2. UPDATE it to completed (with items_processed and summary) or failed (with error) at the end
3. Link data rows back to the run via a run_id foreign key

Phase 5: Sample Run

Run the agent locally before deploying. Expect the first run to fail — wrong API routes, malformed responses, and bad data are normal. The goal is to catch these issues before deploying a broken agent.

1. Install dependencies — npm install in the agent directory
2. Run the script — SAPIOM_API_KEY=... node index.js
3. Check for errors — fix any crashes (wrong endpoints, auth failures, JSON parse errors)
4. Re-run until the script completes without errors

If the agent processes a list of items, the sample run will naturally cover the full set — don't artificially limit it. The point is to verify the pipeline works end-to-end with real data.

Phase 6: Quality Check

A successful run doesn't mean the output is correct. Write a throwaway script that reads the agent's persisted results and checks whether the data actually makes sense.

What to check:

• Are LLM outputs grounded? — If the agent uses search results to inform analysis, do the search results actually relate to the query? Generic or unrelated results mean the search queries need reworking or the search step should be removed.
• Is structured data well-formed? — If the agent extracts JSON from LLM responses, are the fields populated with real content or hallucinated/empty values?
• Are external API results relevant? — Read a few records and verify that data from external services (search, scrape, enrichment) actually matches what was requested. Bad upstream data will silently produce bad downstream analysis.

If the quality check reveals issues:

1. Fix the agent script — adjust prompts, remove unreliable data sources, fix queries
2. Clear stale data — drop and recreate tables if the schema changed, or delete bad rows
3. Re-run from Phase 5

Repeat until the output holds up to inspection. Only then proceed to deploy.

Phase 7: Deploy

STOP — Before deploying, verify: 1. Phase 5 (Sample Run) completed — the script runs without errors locally 2. Phase 6 (Quality Check) passed — the output data is correct and makes sense If either is incomplete, go back. Deploying a broken agent wastes time — it will fail the same way in the cloud.

Read references/deploy.md for the full API reference. Deploy is non-blocking — it returns immediately with status: "building". Poll GET /v1/jobs/{name} every 10–15 seconds until "deployed" or "failed" (typically 60–120s).

Quick summary:

• Scheduled job (most common): ZIP + POST /v1/jobs?name={name}&schedule={cron} → poll until deployed
• Batch job: same deploy, omit schedule, trigger via POST /v1/jobs/{name}/executions
• Orchestrator + workers: scheduled job triggers batch job executions

Cron schedule selection:

Default to every 2 hours (0 */2 * * *) unless the user specifies otherwise.

Phase 8: Verify

After deploy completes:

1. Confirm job status is deployed in the response
2. Report to user: job name, cron schedule, services used
3. Remind user to set SAPIOM_API_KEY env var if needed

Troubleshooting

safeFetch is not a function

Cause: Wrong import or missing @sapiom/fetch dependency. Fix: Ensure package.json has "@sapiom/fetch": "^0.3.0" and code uses const {createFetch} = require("@sapiom/fetch").

HTTP 402 Payment Required

Cause: Missing or invalid SAPIOM_API_KEY. Fix: Check that the env var is set and the key is valid. If running locally: SAPIOM_API_KEY=sk-... node index.js. If deployed: pass via envs in the deploy body.

Scrape returns empty/wrong content

Cause: Firecrawl got blocked, or the page requires JavaScript rendering. Fix: Try adding waitFor: 3000 to the scrape options. Some sites block automated requests — test the URL manually first.

LLM returns garbage or hallucinated data

Cause: Input content too long, prompt too vague, or wrong model. Fix: Truncate input (.slice(0, 8000)), be more specific in the system prompt, or switch to a more capable model (e.g., anthropic/claude-opus-4.6).

Deploy returns 502

Cause: Bad package.json — missing main, scripts.start, or invalid JSON. Fix: Ensure package.json has "main": "index.js" and "scripts": {"start": "node index.js"}.

Agent runs but produces no output

Cause: Env vars missing in deployed environment. Fix: Pass SAPIOM_API_KEY via the envs field when deploying. The job needs it at runtime to call other gateways.

Checklist

• Parsed user intent (what, how often, what output, compute pattern)
• Selected capabilities (read only needed references)
• Got API key (MCP tool, manual, or user-provided)
• Wrote script from template + references
• Sample run completes without errors
• Quality checked output data for correctness
• Deployed via Sapiom (ZIP upload flow)
• Verified DEPLOYED status
• Reported results to user