Indie agent week: pseudocode editors, token ledgers, and repo swarms (2026)

Indie agent builders spent this week on the parts of the stack that chat windows still hide: how you state intent, how much a session costs, whether rules actually fire, and what happens when agents start opening PRs without you.

Window: roughly 14–21 August 2026. Focus is solo and small-team shipping — open tools, Show HN drops, experiments — not frontier lab model cards.

How you talk to the agent

Huzzah: keep intent as pseudocode, not chat debris

Daniel Vaughn is tired of writing long English prompts that vanish after the model rewrites the file. His experiment, Huzzah, flips the defaults: prompts become pseudocode, declarative, and persistent. You edit a .hz file; on save, Huzzah diffs it and regenerates the real code from that intent record.1

The pitch is control and documentation at the same time. The human-authored pseudocode stays as the authority for what the software should do; the generated source is a projection. Vaughn is explicit about the limits: better for greenfield work, awkward for cross-file messes, and still experimental. Source and setup live in danielvaughn/hz. The Show HN thread hit hard this week (285 points).2

Vomit: stop paying Claude to write worse English

Zach Ahn hates Claude 5's "explanations" enough to build an art project around them. Vomit hooks Claude Code, buffers the model's message, rewrites it through a local LLM (he uses GPT-OSS 20B), and shows the cleaned text instead.3

It does not touch Claude's tool calls or files — only the display path — and Ahn is frank that the local model can hallucinate and slow the session down. The useful idea is architectural: stop spending frontier tokens on tone instructions the model ignores; put a small model on the prose and keep the big one on the work. Repo: zachahn/vomit.4

What the sessions actually cost

Frugal Tokens: one local ledger for five coding agents

dpc94 / dpclark4 shipped Frugal Tokens after staring at his own session bills. It is a local, read-only viewer over OpenCode, Claude Code, PI, Codex, and Cursor sessions: spend, cache reuse, miss cost, overlapping sessions, model mix, and a per-session explorer that jumps to the call where a cache miss happened.5

You run it with Deno; the public demo is scrubbed but shows the shape of the UI at demo.frugaltokens.com. Source: dpclark4/frugal-tokens. If you run multi-agent days and only check provider dashboards after the fact, this is the week's practical instrumentation drop.6

Rules and memory that outlive a chat

Rules as Programs: constraints that watch the agent

Yuntian Deng and ProgramAsWeights open-sourced Rules as Programs: each agent rule becomes a small executable check (Python, optionally with a local neural "fuzzy" program) that runs beside the coding agent, scores whether the rule was met, and surfaces findings in a menu-bar inbox. First integration is Cursor hooks; v1 is deliberately non-blocking — audit and observability, not automatic rewrites.7

That is the opposite of stuffing more prose into AGENTS.md and hoping. The rule gets a life of its own after the model has already moved on.

ai-memory: hand off work between vendor CLIs

Fabio Akita's ai-memory kept climbing this week (~3.7k stars on the weekly trend board). The job is blunt: quit Claude Code mid-task, open Codex in the same directory, and keep architecture notes, failed approaches, and open questions without re-briefing the new agent. Hooks + MCP across major coding CLIs; long-term store is the product.8

If your week is multi-tool rather than multi-agent, this is the glue piece.

Skills and swarms

/wayfinder: plan through the fog of war

Matt Pocock released /wayfinder for the part of agent work that still burns humans: planning when the end state is unclear. An orchestrator layer splits research, prototypes, and "grilling" sessions, keeps a shared map of decisions, and hands child sessions a focused ticket so you are not babysitting context windows all afternoon.9

It sits in his broader AI Skills for Real Engineers collection. Useful when you want AFK agents overnight but the plan itself is still half fog.

diagram-design: stop shipping Mermaid slop

Cathryn Lavery's diagram-design skill was the week's GitHub noise machine — tens of thousands of stars on the weekly board. Thirty-nine editorial diagram types as self-contained HTML/SVG for Claude Code, Codex, Factory Droid, and Pi: architecture, loops, Sankey, Wardley, journey maps, and more, with a taste rule that prefers deletion over decoration.10

Not a runtime agent, but it is agent-native craft infrastructure: the thing you install so your coding agent stops drawing the same rounded boxes.

Your repository is your swarm

Nick Trevino wrote up wiring headless Google Antigravity into self-hosted Gitea: label an issue → agent runs; scheduled jobs that open improvement PRs twice a day; agents answering review comments; then a merge-queue bottleneck he asked the agents to build for him. He still reviews and merges by hand. He calls the late-night loop bingeineering.11

The post is an experiment log, not a product launch — and that is why it belongs here. It shows the next failure mode after "agent can code": PR flood, CI thrash, and the need for swarm-level architecture (event storms today, maybe reconciliation loops later).

Harnesses and sandboxes

seed: one file, one tool, grow the rest

Vivek Haldar's seed is almost aggressively small. Frozen layer: seed.py — a loop, a language model, a single exec bash tool, and a system prompt the agent may rewrite. Everything else (tools, memory, skills) has to be grown into a self/ directory across sessions. Plant with uvx, talk, let the individual diverge.12

It is a philosophy demo as much as a product: frameworks optional; the agent owns its own scaffolding.

TakoVM: job queue + gVisor for untrusted agent code

Tako Research shipped Tako VM as a self-hosted execution layer: run untrusted (often model-generated) Python in gVisor-isolated containers with a built-in job queue, workers, retries, execution history, and replay — the bits people usually bolt on after e2b-style sandboxes. Roadmap language points at durable per-agent filesystems; today each container is single-use.13

If your agent needs to execute code you would not run on the laptop, this is the indie infra drop of the week. GitHub: Tako-Research/TakoVM.

Also shipping

Donkey Cut — free, open CapCut-style editor with chat-driven edits and generation; Mac companion can attach Codex or Claude Code logins for the assistant. Solo-scale OSS, not a lab.14
Mole — deep-research agent as a static Go binary: hard USD budget ledger, claims tied to verified source quotes, MCP so a coding agent can drive it, local data stays local.15

Worth trying first

If you only install two things from this week: Frugal Tokens if you already live in coding agents and do not know where the money goes, and Huzzah or Vomit if the pain is the interface — either intent that disappears into chat, or prose you cannot stand reading. For multi-agent CI dreams, read Trevino's swarm post before you turn on unsupervised PR factories.

References

1
Huzzah — Daniel Vaughn
danielvaughn.dev
2
Show HN: Huzzah
news.ycombinator.com
3
How to fix Claude 5's token vomit — Zach Ahn
zachahn.com
4
zachahn/vomit
github.com
5
Show HN: Frugal Tokens
news.ycombinator.com
6
frugal-tokens README
github.com
7
Rules as Programs README
github.com
8
akitaonrails/ai-memory
github.com
9
The /wayfinder Skill — Latent Space
latent.space
10
cathrynlavery/diagram-design
github.com
11
Your Repository is Your Swarm — Nick Trevino
nicktrevino.com
12
vivekhaldar/seed
github.com
13
Tako VM
tako-research.github.io
14
DonkeyCut/Donkey
github.com
15
lajosdeme/mole
github.com