Open Source AI Harness Profiler – discover where tf your tokens are going

Writing the post from only the facts you provided—no invented versions, dates, or figures.Rekon, published open source under the TryRekon organization on GitHub as TryRekon/Rekon, is positioned as an AI harness profiler whose stated job is to show where tokens are spent in agent and harness runs. The listing reached Hacker News under the AI Agents track as “Open Source AI Harness Profiler – discover where tf your tokens are going,” with the discussion thread at news.ycombinator.com/item?id=48998468. At the time of the summary, that thread sat at 1 point and 1 comment, so the signal is early and thin on public reaction rather than a proven community consensus.

Technically, a harness profiler sits outside the model itself and instruments the layer that drives calls, tools, loops, and intermediate steps. Rekon’s pitch is token attribution: instead of treating total usage as a single bill line, it aims to surface which parts of the harness consume tokens. Without published architecture notes, benchmarks, or version numbers in the source material, the product mechanics to treat as given are only that name, open-source distribution on GitHub, and the token-path focus in the title—not any claimed overhead, supported frameworks, or measurement methodology.

For engineers building agents, that focus maps to a daily cost and latency problem: multi-step harnesses burn tokens on retries, tool schemas, context stuffing, and redundant planning, not just the final answer. A profiler that can point at those sinks is more useful than a monthly invoice when you are tuning prompts, tool surfaces, or loop limits. Builders who already log raw token totals still lack a per-harness breakdown; Rekon is aiming at that gap if the implementation can attach spend to concrete stages of a run.

In market terms, Rekon enters a crowded agent-tooling space where open source is often the entry point for observability before commercial APM-style products lock in. Competing pressure comes from vendor dashboards, framework-native tracing, and ad hoc log grepping rather than from a single dominant open profiler brand called out in the source material. The HN AI Agents placement and single-comment thread mark it as a discovery-stage project: low social proof so far, open repo as the evaluation surface, and no independent adoption figures in the facts provided.

Practical takeaway: if you run multi-call agent harnesses and cannot explain token spend by step, clone TryRekon/Rekon from GitHub and test it against one real workflow before trusting it in production. Watch for whether the project documents how it attributes tokens, which runtimes it supports, and whether the HN thread grows beyond the current 1 point and 1 comment—those are the next real signals that the tool is usable rather than just well titled.