Week of 21 Sep 2026¶

Only part of this week's discussion is summarised so far. Dates are approximate.

What happened¶

In this discussion, the lead architect mentors a developer on structuring AI workflows using the Machine World agent harness instead of relying on standard manual chat interfaces. The conversation begins by analyzing an experimental UI built with Claude, pointing out that unchecked context growth degrades model performance and wastes tokens. The architect introduces a [doctrine term]-inspired Cessation Model Context Protocol (MCP) server that applies decay functions to prune obsolete intermediate reasoning steps once decisions are made, achieving 30x to 40x token efficiency. A practical pilot is explored around automating tuition teacher workflows, encompassing worksheet (tuit) generation, dynamic calendar sync, and transport scheduling via APIs. The speaker emphasizes developing self-improving agent loops attached to a lightweight vanilla harness rather than building static, hard-coded applications. They also evaluate hybrid physical-digital agents capable of orchestrating real-world logistics and agent-to-agent tokenized commerce through payment infrastructure like Stripe. Finally, the team plans to demonstrate these capabilities via episodic narrative video showcases that illustrate how Machine World, Inner Realm, and Marga Sakaccha converge into an AI-native operational system.

Key points¶

Said to exist · An MCP server modeled on the [doctrine term] causal doctrine has been developed to handle memory decay and context cessation.
Said to exist · Speaker 2 built an automated video generation pipeline using Claude to handle screen recording, automatic captions, and intro/outro voice stitching.
Planned · Provide end users a default vanilla harness on Machine World where task-specific AI agents can be dynamically loaded and attached.
Planned · Showcase the depth of Machine World, Inner Realm, and Marga Sakaccha together by creating story-driven visual episodes demonstrating real workflow automation.

Ideas and open questions¶

Idea · Use specialized decision-making models (such as Jett or open-source decision engines) rather than expensive general LLMs to make conditional choices without paying high output token costs.
Idea · Integrate third-party transport networks and local services via APIs into calendar agents to schedule commutes and pick-ups automatically.
Open · Local Sinhala AI language modeling research remains underfunded and underdeveloped, forcing reliance on external multilingual models connected via tools like Whisper.

All details (4 topics, 2 more points)

Context Pruning via Cessation memory MCP and [doctrine term] Logic: Speaker 1 explains that preserving intermediate decision rationale causes exponential context bloat and intent drift. By modeling context decay on the Buddhist [doctrine term] system where causes cease upon producing their fruit (cessation), an MCP decay function drops transient memories, yielding 30x to 40x token savings.
Machine World Harness and Agent Pipelines: The team discusses moving beyond isolated chat interfaces by deploying a neutral 'vanilla harness' in Machine World. This harness connects specialized agent pipelines, lightweight decision-making models, and self-improving agentic loops that calibrate against user intent.
Tuition Teacher Workflow Automation Pilot: To validate the system, the developer examines automating repetitive teacher tasks like tuit generation, syllabus alignment, calendar tracking, and bus or transport synchronization. The architecture targets relieving cognitive friction so human educators can focus on deep conceptual transfer.
Physical Agents, Tokenized Commerce, and Ecosystem Integration: Speaker 1 highlights the long-term vision of hybrid agents operating physical and digital tasks, interacting via agent-to-agent token payments supported by gateways like Stripe. The resulting workflows will be showcased through narrative video episodes connecting Machine World, Inner Realm, and Marga Sakaccha.
Decided · The Cessation context decay function achieves an estimated 30x to 40x token reduction by transmitting only essential memory into inference thinking stages.
Decided · Agent-to-agent transactions using tokens can leverage emerging infrastructure such as Stripe's agent transaction capabilities.