Indie agent builds, Sept 14–21: decision scorers, marker-verified plans, and one CLI grammar

Three of this week's releases hand a narrow decision to a small scoring model, two turn a long agent run into evidence you can read afterwards, and two more widen what an agent can reach without holding every tool definition in its context.

What shipped. Cua released CUA-S1-FORMS, the first of its System One models for computer use. Given a document's structured elements and values, the model decides per element whether to use the value, check a box, click, or skip — a fixed set of choices, scored in one pass. It carries 706,000 parameters and a 2.8 MB checkpoint, and its first training run took under 30 minutes on synthetic data. 1

Why it matters. Cua built the model to find out how far a specialist can go inside the space between a brittle script and a general agent loop. Measured against hosted Jev — the TypeSafe model that answers typed questions with probabilities instead of generating text 2 — the specialist scored 99.7% against 83.6% across the whole decision set, and left already-filled fields alone correctly 100% of the time against 74%. Scoring a form locally took 7–9 ms against 260–280 ms per hosted call, though Cua says the two samples measure different things. The synthetic data generator, training, evaluation and the driver integration are in the repo under MIT. 3 It appeared on Show HN on September 19, 2026. 1

Try it. The work sits in libs/cua-s1 in the cua repository, next to the driver that executes the actions, so the quickest test is to point the existing form workflow at the small model and compare it with whatever you call today. 3

What shipped. Openlayer published jevals, a Python library that rebuilds standard agent eval metrics on decision models. One evaluate() call takes the messages and tool schemas you sent to the model and returns tool choice, groundedness, scope, answer relevance, completeness, indirect-injection and PHI results together. In the project's quickstart the nine checks come back as a single request: 1,423 tokens, $0.00006, half a second. 4 Backends resolve from the environment — Jev through TypeSafe or Vercel's AI Gateway, Kev or Laya running locally on a Mac, or an emulated chat LLM when that is all you have. 4 It appeared on Show HN on September 21, 2026. 5

Why it matters. Evals usually run on a sample of traffic because the judge is the expensive part, and a judge that scores the same trace differently on a rerun makes a shaky test suite. LangChain replayed one agent's traces through Jev and three frontier judges and reported quality-score variance 92x to 913x lower for Jev, at 0.44 s and $0.00035 a call, against $28.17 for Claude over the same set. 2 Jev answers three question shapes — pick one, score against a rubric, or return a yes/no probability — which covers most of what an LLM judge is asked. 2 Because an eval only depends on the sample it is given, the same class runs offline, as a monitor on production traces, and as a gate inside the agent loop. 4

Try it. Run pip install jevals, set AI_GATEWAY_API_KEY for Jev through Vercel, or pip install "jevals[laya]" for a fully local model on Apple Silicon, then start from examples/quickstart.py. 4

What shipped. Jeff is a read-only Go CLI that checks source files against semantic rules and prints a pass or a violation. Version 0.1.0 ships four rules: failure-path integrity, side-effect transparency, within-file duplication, and naming that contradicts behavior. Rules live in YAML fragments, configuration in jeff.toml, and output is text or JSON with exit codes — 0 for all checks passed, 1 for a violation, 2 for a usage or inconclusive result. 6 It appeared on Show HN on September 18, 2026. 7

Why it matters. The exit codes are the point: a semantic check that returns a verdict rather than a paragraph can sit in CI next to your linter. The privacy boundary is declared too — only recognized source files are sent, and certificates, keys, .env files, YAML, JSON, TOML, Terraform variables and Markdown stay local unless a rule explicitly allows them. 6 Jeff needs a TypeSafe credential, either TYPESAFE_API_KEY or a stored one in the OS keyring, so it costs a signup before it costs anything else. 6

Try it. Download the release archive and checksums.txt for your platform, or build it with go build -o ~/.local/bin/jeff ./cmd/jeff, then run jeff check src in a repo with a jeff.toml. 6

What shipped. Ordewell turns one goal into an ordered plan of coding-agent tasks, each carrying its own runner, model, thinking effort and mode. The plan is a file: you can rewire dependencies, swap a model, or delete a task before anything runs. Execution then spawns a real session per task, respects the graph, and hands each task its predecessors' results. Claude Code, Codex and OpenCode can serve as planner and runner on the subscription you already hold, with 25 API providers as the alternative, and the planner itself is read-only — it researches the repo and asks questions when the goal is vague. Apache-2.0, Node 20 or newer. 8 It appeared on Show HN on September 15, 2026. 9

Why it matters. A task counts as complete only when its completion marker shows up in that session's output; the exit code is kept as diagnostic evidence, and a task that ends without its marker fails loudly. That keeps the verdict in the transcript a person can open instead of in a model's summary of its own work, which matters the moment a plan runs overnight. 8

Try it. npm install -g ordewell and run ordewell for the terminal UI, or skip the UI entirely with ordewell plan --goal "Add rate limiting to the public API" && ordewell run. The TUI needs tmux; a VS Code extension ships the same core. 8

What shipped. Panel is a local research workspace where chat, files, PDFs and Jupyter notebooks sit in one dock, and the agent can write a new pane when no built-in view can show the thing you asked about. Modules carry typed definitions for inputs, outputs and intermediates, so a long-running job can report progress and its results can be validated as it goes; a data abstraction layer maps a URI to memory or to a local file. MIT, Node 22.18 or newer, and Claude Code signed in locally. 10 It appeared on Show HN on September 15, 2026. 11

Why it matters. A chat transcript is a thin window onto an agent that is reading papers, running notebooks and plotting results, and the usual fix is to keep switching applications. Panel puts the switching inside one workspace and lets the agent add the view it needs — a PDB viewer or a SQLite visualizer — instead of describing the output in prose. Its README states the current limits plainly: Claude Code is the only fully supported agent, modules start by asking the chat, and the hypothesis modules have no view of their own yet. 10

Try it. pnpm install && uv sync && pnpm start, then open http://localhost:4173. Conversations and workspaces live in ~/Panel, outside the repository, so re-cloning does not lose them. 10

What shipped. aclif is a framework for building command-line tools that agents drive, and npm install -g @aclif/core installs one command grammar, one JSON envelope and one error vocabulary across providers. A command's definition loads only when the agent asks for it: --schema, --examples and --shape return before the command runs, need no credentials and spend no quota. Alias sets and a tenant catalog give the same record the same name across platforms. Every command declares its mutability, blast radius, reversibility and idempotency, honours --dry-run and --confirm, and writes an audit line per run. Salesforce, ServiceNow, DocuSign and Agentforce are built in; Node 22 or newer. 12 It appeared on Show HN on September 17, 2026. 13

Why it matters. An MCP server publishes a fixed tool list, and every entry on that list occupies the agent's context on every turn, which forces the server's author to trade coverage for cost when the list is written. Loading a command definition on demand keeps the whole API reachable. The same command classes then run in three places — spawned by the agent, called by a host application, or served by one long-lived gateway that holds the credentials per request — so the choice of where secrets live stays yours. 12

Try it. aclif discover --json lists providers and whether their credentials are configured, aclif learn salesforce --json briefs the agent on fields and query syntax, and aclif salesforce data query --query "SELECT Id FROM Account LIMIT 3" --dry-run validates a real call without touching the instance. 12

What shipped. skillbay.sh is a marketplace of agent skills curated by a human. Every listing is reviewed by a person and shows the same prompt answered with and without the skill, so you can see what the skill changes before you install it. Listings are free and span coding, writing, research, data, devops, design, legal and personal productivity; a wanted board collects requests, and skills install over MCP. The newest entries posted on September 17 and 18 include live San Francisco tennis court availability and an inference-engineering skill. 14 It appeared on Show HN on September 17, 2026. 15

Why it matters. A skill is a lot of prompt in every request and it quietly changes how the agent behaves on unrelated work. The before-and-after prompt is the part of a listing worth reading, because it shows the difference the skill makes on a task rather than describing the skill's intent. 14

Try it. Browse skillbay.sh, or point an MCP client at skillbay.sh/mcp so the agent can search the shelf itself. 14

The seven share one habit: each one moves a judgment into a file, log or schema that outlives the conversation that produced it. Take the one that matches the check you still do by hand. This roundup runs weekly, on Mondays.