5 Must‑Watch Tech Projects This Week: Agent Engineering, Rust CLI Tools, and the Dawn of an Agent Economy

1. mattpocock/skills – Claude Code engineering skill set

Repository: mattpocock/skills (Shell)

Problem: AI coding agents often generate code with poor architecture, inconsistent naming, and terminology that does not match the project's domain.

Solution: Introduces slash commands that shape Claude Code output: /grill-me – repeatedly asks the agent to clarify intent before code generation. CONTEXT.md – a project‑level glossary that forces the agent to use domain‑specific language. /caveman – compresses the agent’s output, claiming a 75 % token reduction while preserving technical accuracy. /diagnose – enforces a disciplined debugging loop (reproduce → minimize → hypothesize → verify → fix).

The author emphasizes that controlling agent behavior is more valuable than merely adding capabilities.

2. Hmbown/DeepSeek‑TUI – Rust‑based terminal programming agent

Repository: Hmbown/DeepSeek‑TUI (Rust, 94.9 % of code)

Problem: Claude Code ties users to Anthropic’s API, limiting model choice.

Solution: Provides a pure‑terminal programming agent with three operation modes and OS‑level sandboxing:

Plan – read‑only mode, no code execution.

Agent – step‑by‑step confirmation before each action.

YOLO – fully automated execution.

The modes correspond to Claude Code’s --dangerously-skip-permissions flag but can be toggled at runtime. The Rust implementation adds macOS Seatbelt and Linux Landlock sandboxing, improving security compared with most peer tools.

3. rohitg00/agentmemory – Persistent memory layer for AI coding agents

Repository: rohitg00/agentmemory (TypeScript)

Problem: CLAUDE.md’s memory is volatile, limited by line count, becomes stale, and requires loading the entire context on each run.

Solution: Replaces CLAUDE.md with a searchable database that automatically captures tool calls and session summaries. On restart, only relevant memory fragments are injected.

Hybrid search combines BM25 keyword matching, vector embeddings, and a knowledge‑graph via Reciprocal Rank Fusion.

Four‑tier memory architecture: work → episodic → semantic → program memory, with automatic decay and forgetting.

Supports 51 MCP tools and integrates with Claude Code, Cursor, Gemini CLI, etc.

On the LongMemEval‑S benchmark the system achieves 95.2 % recall@5, and the author claims a 92 % token reduction compared with the original volatile memory.

4. Zed 1.0 – GPU‑accelerated open‑source editor

Repository: zed‑industries/zed (Rust)

Release date: 2026‑04‑29

Problem: Existing editors (VS Code, Cursor) rely on Electron/Web UI, imposing performance ceilings.

Solution: Implements a custom GPU UI engine (GPUI) that renders directly on the GPU, delivering fast startup and responsive interaction. Version 1.0 adds AI‑focused features:

Parallel Agents – multiple agents run concurrently without interfering with each other.

Agent Client Protocol (ACP) – a native protocol that lets external agents (Claude Code, Codex CLI, Cursor, etc.) read/write files, navigate code, and invoke tools directly inside the editor.

The editor’s creators are the original authors of Atom and Tree‑sitter, providing deep expertise in editor architecture.

5. pay.sh – Native API payment gateway for agents

Source: Solana Foundation × Google Cloud (launched 2026‑05‑05)

Link: https://pay.sh

Problem: Current agent‑driven API calls follow human workflows—account registration, credit‑card entry, API‑key management—creating barriers for autonomous agents.

Solution: Uses Solana‑based USDC for per‑request billing. An agent only needs a Solana wallet to discover and invoke over 50 providers (including Google Cloud Gemini, BigQuery, Vertex AI). This eliminates account systems, supports micro‑payments, and provides transparent settlement. Context: The same week AWS previewed Amazon Bedrock AgentCore Payments, indicating strong industry momentum for autonomous API consumption.

Summary of the stack

Behavior control: mattpocock/skills – aligns agent output with user intent.

Runtime: DeepSeek‑TUI – runs a full programming agent in the terminal.

Memory layer: agentmemory – preserves context across sessions.

Editor integration: Zed 1.0 – turns the editor into a collaborative agent interface.

Economic layer: pay.sh – enables agents to autonomously pay for resources.