[How-To] GasTown Workflows & 60-Second OpenClaw
Feb 26 Office Hours Recap: How to slash multi-agent token costs, deploy Kimi Claw in 60 seconds, and why "Intent Engineering" is the new standard
This week’s AI Center of Excellence Office Hours tackled the economics and mechanics of advanced agentic orchestration. We explored how to run the token-hungry GasTown framework without going bankrupt, dissected the trade-offs of Kimi’s one-click OpenClaw deployment.
Here is the real alpha from the February 26th session.
1. GasTown: The “Kubernetes of Agent Orchestration”
Stephen Barr gave a masterclass on GasTown, a multi-level orchestration framework for AI agents. Modeled loosely around a Mad Max theme, it utilizes a hierarchy of specialized agents to burn through software engineering tasks in parallel.
The Architecture:
The Mayor: The orchestrator agent that sequences tasks and dispatches work.
Pole Cats (Workers): The agents that actually execute the code specs.
The Refinery: An agent dedicated to observing the merge queue and resolving Git conflicts from parallel work streams.
The Deacon & The Witness: Agents that monitor the health and activity of the system.
The Problem: If you run GasTown on default settings using Claude Opus, you will bankrupt yourself with token costs.
The Hack: Stephen demonstrated how to configure GasTown via the open-code CLI to use a mix of models. By assigning Claude Sonnet to the “Mayor” role and routing the “Worker” tasks to highly capable, cheaper models like GLM5 or Kimi K2.5, you can cut execution costs from $5.00 down to 45 cents per job.
Task Persistence: Stan Huseletov highlighted GasTown’s unique advantage: it does not care about “sessions”. It relies entirely on tasks (called “Beads”) and the Git work tree. Beads are backed by DOL, a SQL database that is versioned like Git, meaning you never lose your place if a session dies.
Resources:
Stephen’s Deck: GasTown Presentation (Made with Kimi Slides)
![[Deep Dive] Gastown](https://substackcdn.com/image/fetch/$s_!tPGN!,w_140,h_140,c_fill,f_webp,q_auto:good,fl_progressive:steep,g_auto/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fae6ba179-e3c1-4584-8fa7-b6bc5e815886_2752x1536.png)
2. The 1-Minute OpenClaw (Kimi Claw)
Last week, we debated the risks of hosting OpenClaw locally. This week, Leonardo Gonzalez demoed Kimi Claw, a web-based, one-click OpenClaw deployment.
The Trade-offs:
Speed: Deployment takes literally 60 seconds from the Kimi interface.
Cost: It requires the “Algretto” $40/month coding plan, which provides a fixed-cost API key for use.
Capability: It features prompt-based self-management (it can install skills from Claude Hub directly) and persistent memory, treating past conversations as fragments that must not be deleted.
The Catch: It is a black box. Unlike a terminal-based local deployment, you cannot look under the hood to manually curate the SQLite memory databases or debug custom shell scripts easily.
Security Insight: Leonardo emphasized a strict security posture: never give a cloud-deployed agent access to your unmetered enterprise API keys. Rely exclusively on the fixed-cost $40/mo Kimmy/GLM keys to sandbox your financial risk.
Resource: Kimi Bot
3. Alternative Models Now on AWS Bedrock
If you are still exclusively relying on Anthropic, you are missing the Pareto frontier of performance-to-cost. Stephen Barr announced that GLM5, Kimi K2.5, and Minimax are all now officially supported on AWS Bedrock.
GLM5: Half the size of top-tier models, explicitly trained for agentic tool use. Read the GLM-5 Blog.
Minimax m25: Half the size of GLM, offering incredible speed and cost-efficiency for high-volume tasks.
Kimi K2.5: Highly recommended for complex reasoning tasks as the “driver” of your OpenClaw setup.
Resource: AWS Bedrock Adds Support for Six Open Weights Models
4. Is Prompt Engineering Useless?
To manage context and agent focus, I uses a monetary tiering hack: he assigns tasks a strict value (e.g., $10, $100, or $1,000) so the orchestration swarms always know what the actual priority is without getting lost in verbose prompts.
For those who still struggle with vague prompts, I shared a custom tool he built to automatically analyze and mistake-proof user inputs before sending them to an LLM.
Tool: Perfect Prompt Analyzer
Counter-Resources (For the Prompt Purists): Anthropic Context Engineering and OpenAI Prompt Guides
5. Rapid Upskilling: Learners Lens Courses
Once again, the team at Learners Lens proved how fast AI can generate curriculum. Based purely on this week’s Office Hours agenda, they generated three comprehensive video-based courses in under 10 minutes:
Check out The AI First Show on YouTube for short video recaps of today's live demos, and subscribe so you never miss a technical deep dive.