Bivor is a free, MIT-licensed macOS desktop app that wraps the pi coding agent in an orchestration layer: it runs each chat as its own isolated pi process, gives every task a git worktree and branch, adds per-tool guardrails and turn/cost budgets, and snapshots the workspace before every prompt so you can roll back. It reads the same ~/.pi/agent/ session and config files as the pi CLI, so terminal and desktop sessions interoperate. It launched as v0.1.0 on 2026-08-15 and shipped v0.1.1 three days later. ...

A2A routing agents decide which agent configures, owns, or continues an in-flight task once execution has already begun. The Sprix SAGE Router is a dependency-free Python reference implementation of that policy: it compares SELF, COLLABORATE, and HANDOFF under permission, budget, and deadline constraints, and returns an auditable trace. Its own trajectory replay shows progress-awareness mainly buys restraint — wasted work falls from 0.104 to 0.059 and the switch rate from 80.1% to 58.4%, while utility stays inside overlapping error bars. ...

Nopus is a deterministic prose checker for coding agents: it measures a finished response with packaged English word data, and if several signals cross their thresholds it asks the same agent for exactly one clearer rewrite. There is no LLM judge, no network call, and no retry loop. That is the whole product. nopus (npm @syzom/nopus, MIT, by GitHub user Vistyy) shipped on 2026-08-15 and reached its current feature set in five days across 14 commits. It supports exactly three hosts — Pi, Claude Code and Codex — and it does one thing: after an assistant finishes answering, it looks at the prose and decides, reproducibly, whether that prose was too hard to read. If so, the agent is told once to say it more plainly. ...

Walking dead state detection is the practice of proving that a game can reach a state where it still accepts input, still renders, still answers every command — and victory has already become impossible. The player does not know it. The game does not crash. Detection means finding those states from the game’s own logic, or from an agent that explores its way into them. What Is a Walking-Dead State, Exactly? The term is design criticism, not engineering vocabulary, and it starts with the player rather than the code. Jimmy Maher’s The 14 Deadly Sins of Graphic Adventure Design named the walking-dead syndrome: the human continues playing a game that has, unbeknownst to them, been rendered unwinnable. His worked example is Uninvited (1986). Walk past the mailbox in the Front Yard without looking inside it, step through the front door, and the door locks behind you forever. The mailbox held an item you now cannot get. As Maher puts it, “as soon as you go inside, you become a walking dead.” ...

OJCP (Open Job Context Protocol) is a draft open standard that lets AI agents find, read and act on job postings through MCP tools instead of scraping career pages. A provider publishes a manifest at /.well-known/ojcp.json, implements the required search_jobs tool plus five recommended ones, and normalizes its apply paths. Current status: a living v0.2 draft with 80 GitHub stars. Two disambiguations before anything else. First, this OJCP is not Ontario Job Creation Partnerships, the Canadian government program that owns the acronym in most search results; in this article OJCP means the agent-facing job data protocol authored by Recruitics CTO Austin Anderson and governed by a steering committee that includes Workday, CrossCountry Healthcare, Hiring.cafe, aiApply, scale.jobs, LoopCV and Tink. Second, OJCP is not a rival to the Model Context Protocol. The specification says it outright: “OJCP is not a competing protocol. It is a vertical application of MCP for the job data domain.” MCP defines how agents call tools; OJCP defines which tools exist for job search, application and identity verification. ...

The AutoDesign agent is an open-source design agent that turns one research paper into four editable artifacts — a poster, a slide deck, a webpage, and a narrated captioned video. Its authors report 78.32 on their own PosterBench benchmark against 70.87 for Anthropic’s closed-source Claude Design, with model weights frozen the entire time. That last clause is the part worth your attention. AutoDesign did not win by training a better model. It won by optimizing the system around a fixed model — the scaffold, the tools, the critics, the loop control — and then publishing that scaffold so it can be attached to Codex, Claude Code, Kimi Code, OpenCode, or DeepSeek Harness. The leaderboard number is the marketing. The reusable harness is the actual contribution. ...

Zero-Mem for Pi is a community extension that gives the Pi coding agent long-term memory without spending a single LLM token on memory operations. It keeps every finished session as a raw trace, indexes those traces with local NER and embedding models, and injects a small number of retrieved snippets into later prompts — so the model is called once, for the answer. That is the whole idea, and it is narrower than the marketing line suggests. The architecture comes from the paper Zero-Mem: Zero-Token Memory Operations for LLM Agents (arXiv:2607.29377v1, 31 July 2026, CC BY-NC-SA 4.0), which reports a 57.6% reduction in memory-operation time cost versus the fastest baseline at matched final-answer quality. Two independent Pi ports exist as of this writing: skorotkiewicz/zero-mem (the early adaptation, roughly two stars, README package name pi-zero-mem) and woolcoxm/zero-mem-pi (a self-described “faithful reimplementation”, 38 stars). Both are architectural adaptations of a paper whose reference implementation is still pending peer review, and both say so. ...

Palomar is a public, searchable registry of Lean 4 formalizations whose proofs have been machine-checked and whose statement was reviewed for fidelity to the informal claim. It is jointly incubated by the Lean Focused Research Organization (Lean FRO) and ICARM, announced publicly by Terence Tao on 18 August 2026, and it deliberately performs no human peer review. That last clause is the part most coverage gets wrong, so it is worth stating before anything else: a Palomar entry certifies that a proof typechecks under two independent kernels, that no forbidden axiom was smuggled in, and that an automated reviewer found the Lean statement faithful to the stated informal one. It certifies nothing about whether the result is novel, interesting, or important. ...

A sandboxed agent harness is the layer that gives each AI agent its own isolated computer, routes every outbound request through a policy gateway, and injects credentials at the network boundary so the agent never holds a real key. OneCLI (YC Summer 2026, built by ChartDB Inc.) ships exactly that: one sandboxed agent per employee, all egress forced through a Rust gateway that enforces organization policy before a secret is ever decrypted. ...

Frugal Tokens is a local, read-only Deno dashboard that reads the session transcripts Claude Code, Codex, OpenCode, PI, and Cursor already write to disk, then reports spend, estimated working time, and — its differentiator — the type and dollar cost of every prompt-cache miss. It answers why a bill grew, not just how large it was. That distinction is the whole reason this tool deserves a closer look than its 25 GitHub stars suggest. Coding agent cost tracking in 2026 has become a solved collection problem and an unsolved interpretation problem: every serious tool parses the same local JSONL files, so the only thing left to compete on is what you do with them. Frugal Tokens bets on cache economics. This review tests whether that bet is worth running. ...