Introduction to OpenClaw Optimization

OpenClaw is described as the number one open‑source project in the history of humanity. The speaker argues that every enterprise and software company needs an agent strategy, specifically an OpenClaw strategy, and notes having spent over 200 hours and billions of tokens perfecting their setup.

Key Optimization Techniques

Threading for Memory Management

The biggest and easiest unlock is using “threads” within chat applications such as Telegram, WhatsApp, or Discord. Instead of a single long chat window, separate threads are created for distinct topics—general, CRM, knowledge base, cron updates—by making Telegram groups that include only the user and the OpenClaw bot. This prevents interleaving of topics, keeps each context window focused, and makes it easier for OpenClaw to stay on topic and remember information.

Voice Memos for Input

Holding the microphone icon in Telegram records a voice memo that is sent directly to OpenClaw. This enables asynchronous communication without typing long prompts, saving significant time especially when mobile or driving. Voice memos can be used for questions, task instructions, or coding requests.

Agent‑First Platforms

Here.now Overview

Here.now, the video sponsor, is an agent‑first platform built from the ground up exclusively for agents. Agents can publish any artifact—PDFs, HTML, images—without needing a backend. Setup is as simple as copying instructions into the agent. The service is free to start, with costs only for heavy usage. Published content expires after 24 hours unless an account is claimed, and the front‑end can be edited easily (e.g., switching to a text‑only view).

Model Selection and Management

Multi‑Model Approach

Using a single model for all OpenClaw tasks is discouraged. A wide spectrum of models—frontier and local open‑source—should be employed because different models excel at different tasks.

Assigning Models to Tasks and Threads

• Main chat/orchestration: Sonnet 4.6 (or Opus 4.6) for planning and coordinating sub‑agents.
• Fallback: GPT 5.4.
• Coding & nightly runs: Opus 4.6.
• Non‑frontier tasks: Sonnet.
• Search: Grok.
• Video processing: Gemini 3.1 Pro.
• Deep research: Gemini Deep Research Pro.
• Training pipelines (open‑source): GPT 5.4 extra high.
• Embeddings: Nomic.
• Local models: Quen 3.5.

OpenClaw stores model configurations and fallbacks, allowing users to specify which model to use for a given use case. A new feature lets specific models be assigned to particular threads, delivering faster, cheaper results and optimizing the token budget.

Open‑Source Model Integration

Autonomous systems are explored to identify fine‑tuning opportunities for small local models. For example, a fine‑tuned Quen 3.5 (9 billion parameters) performed email labeling on par with Opus 4.6 at the cost of electricity alone.

Delegation and Sub‑Agents

Benefits of Delegation

Delegating long‑running tasks to sub‑agents prevents the main agent from being blocked, allowing continuous interaction with the primary agent.

Identifying Delegable Tasks

Tasks such as all coding work, search API calls, multi‑step data processing, file operations beyond simple reads, calendar/email actions, knowledge‑base ingestion, and any operation exceeding 10 seconds are prime candidates. Simple conversational replies, quick reads, and status checks are not delegated.

Agentic Harnesses

Delegation can target specific agents like the Cursor Agent CLI or harnesses such as Claude Code. These harnesses take over entire tasks end‑to‑end and return summaries. Early and frequent delegation, using faster/cheaper models for simple sub‑tasks, is recommended.

Prompt Engineering for Multiple Models

Model‑Specific Prompt Challenges

Different models (e.g., Opus 4.6 vs. GPT 5.4) have distinct prompt preferences regarding capitalization, explicit instructions, and negative constraints. A single prompt file optimized for one model may underperform with another.

Managing Multiple Prompt Files

The strategy is to create separate prompt files optimized for each model, download best‑practice guidelines from frontier labs, and instruct OpenClaw to generate optimized versions. Primary prompts reside in the root directory, while fallback prompts are stored in a subdirectory (e.g., /gpt). Scheduled cron jobs keep prompt files synchronized.

Scheduled Tasks (Crons)

Crons automate routine work such as sponsor inbox refreshes, polls, daily backups, syncs, and health checks. Running them during off‑peak hours offloads compute and spreads usage across quota windows (e.g., Anthropic, Codex), preventing quota exhaustion during daytime operations. Example cron jobs include documentation drift checks, prompt quality verification, PII/secrets review, and HubSpot synchronization.

Security Best Practices

Prompt Injection Defense

External data can be “poisoned” with text designed to hijack the model. A two‑layer system is used:
1. Deterministic text sanitation scans for known injection patterns.
2. Frontier scanner (GPT 5.4 or Opus 4.6) reviews any text that bypasses the first layer, assigns a risk score, and quarantines dangerous inputs.

Data Exfiltration Prevention

Outbound data is scanned for secrets or personally identifiable information (PII). Sensitive details such as phone numbers and emails are aggressively redacted in all outputs (Slack, email responses).

Granular Permissions and Approvals

AI agents receive only the permissions they need (e.g., read email but cannot send). Destructive actions require explicit approval.

Runtime Governance

Spending caps, volume limits, and loop detection are enforced to avoid token‑budget exhaustion or recursive loops that could generate large bills.

Logging and Debugging

Comprehensive logging is cheap and essential. Approximately 1 GB of logs covers two months of activity. Instead of debugging directly with the AI, the system can be instructed to review logs, identify warnings or errors, and propose fixes. This approach has resolved issues such as Google O‑break failures and sponsor inbox refresh problems.

System Updates and Subscriptions

OpenClaw releases updates roughly every day or two. A cron job can check for new releases, pull changelogs, summarize changes, and restart the system automatically, keeping security and features current. Using subscriptions (e.g., Anthropic Pro, ChatGPT Plus) via SDKs within OpenClaw is far cheaper than making individual API calls.

Documentation and Version Control

Good documentation—covering agents, souls, identities, tools, heartbeats, memories, sub‑agent policies, PRDs, use‑case workflows, and prompting guides—enhances OpenClaw’s effectiveness. A cron can compare documentation against code to flag gaps. Git tracks code changes, while GitHub stores snapshots in the cloud for backup. Non‑code assets (databases, images, PDFs) are backed up using services like Box, which offers a CLI that agents can invoke.

Testing and Notifications

Tests are written for all code to verify expected outputs (e.g., confirming 2 + 2 = 4). Notification batching reduces noise: low‑priority alerts are summarized every three hours, higher‑priority alerts hourly, and critical issues are delivered immediately.

Development Environment

While OpenClaw can be built via Telegram, using an external development system streamlines iteration. Preferred tools include Cursor, with alternatives such as Codex, Claude Code, or Sublime Text. These editors are optimized for “vibe coding,” offering version control integration and efficient debugging.