Posts

Unreal Agent is an async-first agent harness from Unreal Labs: a Go library, a runner binary, and a Harbor-compatible benchmark runner, MIT licensed, first released on 22 September 2026. When the model issues a tool call, the harness immediately appends an event-log record in the in-progress state, lets the model keep working, and appends the real result when the tool finishes. It is not Epic’s Unreal Engine. It is also not Inworld’s Unreal Engine AI toolkit, which Inworld launched as the “Unreal AI Runtime” — two unrelated products already occupy that name. ...

If you let a coding agent run unattended, the only durable record of what it did is the transcript of its tool calls. Dashboards show you totals; the transcript shows you the sequence. ToolReplay is a small, dependency-free Python CLI that reads one such transcript and answers three narrow questions about it: did the same call ever return two different answers, did the agent repeat a call that could not have changed anything, and did it call a tool it was never granted. It reads a JSON Lines file, and it does the work offline. ...

Selecting an agent-memory system on retrieval quality is measuring half the system. The half that produces confidently-wrong agents in production is the write path — whether a fact is actually recorded, whether an update supersedes the old value, whether a contradiction is flagged or silently averaged, and whether a delete truly removes the fact from every store it touched. Almost nobody publishes how to measure that. This article is a fixed, reproducible evaluation protocol for the write path, and it ships the working parts: a controlled dataset, a generic harness with a documented adapter contract, a scoring sheet, and deterministic tests that prove every promised measure is runnable. It reports no vendor numbers of its own. That is deliberate: the deliverable is the protocol, and it is built so the person who runs it publishes the result — never a vendor benchmark reused as if it were an independent measurement. ...

Magnitude is an Apache-2.0 local inference engine that assesses your hardware from metadata, compiles and tunes its own numerical kernels on that device, then serves an open model on port 10100 and rewrites your coding agent’s config to use it. “Self-optimizing” here means a kernel autotuner under a fixed precision policy, not a learning system. The claims below were re-verified against the live repository, its Apache-2.0 license, the product documentation and the launch thread on 8 October 2026. This is a source-verification review, not a hands-on benchmark: every figure is either the vendor’s own artifact or a named third party’s self-reported measurement. ...

Drawgent is a local server that connects your own coding agent — Claude Code, Codex, or opencode — to a live, persistent Excalidraw canvas that you and the agent edit together. It ships no agent and no API key. It drives the agent you already have installed, in the repository you already have open. That is the whole idea in one paragraph, and it is a genuinely different one. Most “AI whiteboard” tools generate a picture from a prompt and hand you the PNG. Drawgent instead treats the canvas as a two-way work surface: the agent reads what you drew, you read what the agent drew, and — the part almost nobody else ships — individual shapes can be linked to files, line ranges, and symbols in your codebase, then checked in CI so your architecture documentation cannot silently drift. ...

AgentRun (agent.run()) is an open-source TypeScript DSL that describes an agent workflow as a JSON document: a developer-authored graph of typed nodes, with the model invoked only where a node actually needs judgment. You write the workflow once, version it, and run it — instead of letting a prompt loop decide its own next step at runtime. That is the honest one-paragraph answer. What follows is a review of what AgentRun actually ships in beta, where the determinism claim holds up, where it does not, and which teams should care in 2026. ...

AX Check is a free tool by Gauge that scores how far a coding agent gets with your product without human help. It runs 23 checklist items across Clarity, Onboarding, Pricing and Activation, then records three live sessions on DeepSeek, Kimi and Qwen. The grade is provisional and technical-only — the sessions never affect it. What Is AX Check and Who Built It? AX Check (ax-check.com) is an agent usability tester built by Gauge (withgauge.com) and launched on Hacker News on 2026-09-17, where it reached 38 points and 41 comments. Its one-sentence pitch is the whole product: “How far can a coding agent get with your product?” ...

Run N isolated candidate agents in separate git worktrees, capture each trajectory and patch, then rank them with a verifier that returns a continuous logprob score rather than a discrete grade. Terminal-Bench V2 went from 83.1% Pass@1 to 86.5% verifier-selected against a 92.1% oracle — the model had already solved it; selection was the bottleneck. That gap between what the model can do and what you can identify is the entire subject of this guide. The rest of it covers why a plain LLM judge silently fails, how to turn a 20-token score distribution into a calibrated reward, how to rank 8 candidates without paying for 28 pairwise comparisons, what worktrees do not isolate, how many samples are actually worth paying for, and where optimistic best-of-N claims collapse under measurement noise. ...

A requirement ledger is a retrospective requirements record: instead of writing a specification before you prompt an AI agent, you mine the conversation you already had and reconstruct what you actually required. Its evidence trail is the transcript, a test log, and a read-only Git snapshot — and your last correction beats your first prompt. Why do vibe-coded projects lose their requirements? They lose them because the intent was never written down in the first place — it lived in the chat window, and the chat window closes. A prompt is not a requirement document. It is a compressed instruction to a system that fills in every gap with a plausible guess, and the guesses are what you end up shipping. ...

gogent (github.com/tltre/gogent) is an MIT-licensed Go agent framework that deliberately does not write your agent’s logic. It ships the operational layer instead: a daemon that forks each agent as its own OS process, a centralized tool registry with manifest-based authorization, MCP and sandbox routing, credential isolation, and OpenTelemetry spans. It is pre-1.0, single-author, and has 3 stars. That is the whole verdict in one paragraph. The interesting question is not “is gogent popular” — it is not, and the numbers are unambiguous — but whether the layer it occupies is a real gap in the Go ecosystem. Eino, Google’s ADK for Go, Genkit and tRPC-Agent-Go are all in-process libraries that own your agent loop. gogent refuses to, and instead owns the process model around it. That distinction is worth understanding even if you never install it. ...