TL;DR
Be specific, build once, then run on schedule with API-first instructions.
- State the problem clearly and brainstorm with the agent to find the right solution
- Give enough context and mention the tools you use
- Leverage Twin’s built-in tools (research, sheets, slides, email, and more)
- During build mode, monitor your agent and course-correct in the chat if it goes off track
1. How Credits Work
Every credit equals $0.01 of usage. Credits are consumed by model decisions (input/output/cache tokens) and paid tool usage when applicable.First runs cost more
Twin is discovering APIs, handling auth, and shaping reusable tools for the first time. This exploration is credit-intensive.
Later runs cost less
The agent reuses what it already learned — established tools, cached decisions, and proven paths.
Every decision counts, including retries. If an agent retries a failed step, that still consumes credits.
2. Why Scheduled Runs Are Cheaper
Build-style runs discover and design workflows. Run-style executions reuse established tools and instructions — making recurring runs 12x cheaper than builds.Build and refine once
Get the agent working correctly through the chat editor. Iterate until the output is reliable.
Schedule it
Deploy the agent on a recurring schedule. Each scheduled run uses the lightweight run mode.
3. API-First Beats Browser-First
Not all execution paths cost the same. Use this priority order whenever possible:| Priority | Method | Why |
|---|---|---|
| 1 | Built-in tools | Fastest, cheapest, most reliable |
| 2 | APIs | Deterministic and easy to maintain |
| 3 | Scrapers | Good for static pages |
| 4 | Browser automation | Last resort |
APIs
Faster, more deterministic, and easier to maintain. Always preferred when available.
Browser automation
Slower, more brittle, and more expensive. Use only when there’s no API or scraper alternative.
4. Better Instructions = Lower Costs + Better Quality
The quality of your instructions directly impacts both cost and accuracy. Always include:| Detail | What to specify |
|---|---|
| Source app(s) | Exactly where data comes from |
| Filters & constraints | Date range, owners, statuses, thresholds |
| Output format | Table, CSV, JSON, email summary, doc, slide deck |
| Destination | Where the final output should go (Sheets, email, docs, etc.) |
| Definition of done | One sentence that says what success looks like |
Prompt template
Example prompt
Every weekday at 8:00 AM, pull yesterday’s new leads from HubSpot, exclude duplicates by email, append to my “Inbound Leads” Google Sheet, then send me a 5-bullet summary email with total count and top 3 companies.Done when the sheet is updated and the email is sent.
5. Brainstorm and Debug to Keep Agents Efficient
Don’t just build and forget — actively improve your agents over time.Brainstorm tasks
Use Twin to discover high-ROI workflows before building. Ask it what’s possible with the tools you already use.
Debug last run
When costs feel high or results are weak, ask Twin to analyze what happened and suggest improvements.
- “Where did this run spend most credits?”
- “What should I change to reduce cost by 30%?”
6. Keep Agents Focused to Avoid Context Bloat
Long-running agents accumulate context with every message and decision. The more steps an agent takes in a single session, the larger its context window grows — and the more tokens (credits) each subsequent step consumes. To keep costs predictable:- Break complex workflows into multiple focused agents inside a Workspace — the Orchestrator can help you split them
- Avoid open-ended goals like “research everything about X” — set clear boundaries and a definition of done
- Use Run mode for recurring tasks — it starts fresh with the agent’s memory, avoiding context buildup
Built-In Tools
These are available out of the box and unlock a wide range of workflows without any setup:web_search
Fetch current info, API docs, and web answers
scrape
Pull clean content from specific pages
deep_research
Multi-source research with async polling
stock_market
Quotes, historical prices, and market news
send_email
Alerts, summaries, and outbound updates
build_chart
Generate chart images for reports
scrape_platforms
Scrape social, job, map, and marketplace platforms
sheets
Create, read, update, and export Google Sheets
slides
Generate presentation decks (PDF output)
docs
Create and edit formatted Google Docs
generate_media
Create AI images and videos for campaigns and content
Quick Recap
Be specific
Clear goals with filters, formats, and destinations
Build once
Refine until reliable, then schedule
API-first
Prefer APIs and built-in tools over browser automation
Start Building
Apply these tips in the Twin App