Parcha Open-Sources AgentRun: Code Review Drops from $1 to 18 Cents per Check
Agents keep re-reading the same API docs and re-solving the same problems, and you pay for every token. Parcha has released AgentRun under Apache-2.0: after each run, the agent saves a note, a skill, or a ready-made playbook. Its Greptile review bill (one dollar per check) grew from $180 to $6451 a month, while its own reviewer on a cheap model costs 18 cents. For recurring tasks, it makes sense to call the expensive model only for the initial training.
AI-processed from Hacker News AI; edited by Hamidun News
Imagine hiring a very smart employee who forgets everything they learned yesterday. Every morning they re-read the documentation for the same API, re-plan the same task, and make the same mistakes again, and you pay for every hour. This is roughly how many agents in production work today: every repeated run is billed in tokens, even when the task is long familiar. The Parcha team has released AgentRun, an open-source platform that tries to eliminate this repeated work. The project appeared on Hacker News as a Show HN, and all the figures below are claims from the authors themselves, not an independent benchmark.
Where the $1 and 18 cents come from
The authors' motivation is down to earth. Greptile, a code review service, had charged one dollar for each pull request review since March. When agents on the team became more active, the monthly bill rose from $180 to $6451, roughly 36 times higher. Parcha then built its own reviewer on a cheaper model and tuned it on its own PRs. According to the authors, one review now costs 18 cents instead of a dollar. On 30 pull requests the system had never seen, it found 1.6 times as many real bugs as Greptile. The sample is small, so treat this figure as a rough indicator, not as proof.
The second example is merchant classification, meaning determining which category a seller belongs to. First, the expensive Opus model works out how to do the task. Then the accumulated experience is passed to a cheaper model. The cost per merchant fell from $0.89 to $0.12.
How it works
AgentRun does not replace your agent. You choose the harness: Claude Code, Codex, OpenCode, or pi. The platform adds three things around it.
The first is self-learning. After a run, the agent saves what will be useful next time: a note, a skill, or a ready-made playbook. Everything is stored in a separate git folder for the agent.
Human feedback, such as a rating or a message after the first response, becomes lines in a corrections.md file, and every subsequent run receives it in full. When an agent accumulates 20 completed runs (the default), the platform tries to build a playbook from them: a step-by-step document with tool calls, model steps, and quick yes/no decisions with a confidence score.
For example, the model leans toward "approve" with 0.61 confidence, while the playbook threshold is 0.75, so the case goes to manual review.
Before activation, the playbook is checked in code, dry-run, and replayed once on past inputs. It is enabled only if it answers a sufficient number of cases and matches the agent's answers. Complex cases go back to the agent.
The platform rejects secrets in notes and drafts.
The second is durable execution. All runs are stored in a single store. If the server is killed mid-run, the run resumes after restart. If a runner crashes, another runner picks up the work once its 30-second lease expires. Completed steps are not executed again, and a record interrupted by a failure in the middle of a call is marked as interrupted so that you can check it manually.
The third is a control panel. Each tool has a mode: allowed, ask, or denied. A run has a spending limit: when it is reached, the run pauses rather than fails. Cost that the platform cannot calculate is shown as "unknown", not as $0. An agent can be given the computer it needs: a sandbox, a Linux desktop, an Android phone, or a browser.
What you can use right now
The platform is distributed under Apache-2.0. It requires Node 24.22.3 or newer, and runs on macOS or Linux. You can run it on a laptop, in a single container, or as a server with multiple runners. There are instructions for Railway, Render, a VPS, Kubernetes, Coolify, Dokploy, and Docker. If you already work with Claude Code or Codex, you can connect your existing configuration. For Codex and pi, signing in with ChatGPT is enough, with no model key needed. Agents that already run on your backend can simply be observed through the CLI and an OpenTelemetry Collector, without enabling learning.
Ready-made agent templates are included: pull request review by eight independent reviewers, due diligence on a company, a weekly digest of changes at tracked companies, and lead scoring from 0 to 100. Each template can be tested on your own test cases.
Limitations and conclusion
Reflection and independent quality evaluation add token spending, so the savings are not free. Remote connections currently require a server and a Python connector from a single verified checkout. All figures come from Parcha's own tasks, and results on your code may differ.
The practical takeaway is simple. If you have repeated agent tasks such as review, classification, or scoring, it is worth paying a strong model once and transferring the accumulated experience into a playbook for a cheaper one. You can test this idea on your own data right now: AgentRun works well as a ready-made testbed.
- Nov 21, 2026Альтман назвал «ужасной идеей» передачу решений ИИ, а OpenAI бросила 25% инженеров на защиту
- Nov 21, 2026Anthropic shares opened at $111.50, valuing the seven co-founders' stakes at $180 billion
- Nov 21, 2026Phota Labs, Backed by a16z With $5.6M, Teaches AI to Reshoot Bad Photos Without Changing the Face
- Nov 21, 2026Author Makes a Narrative Video on Home LTX 2.3 (22B Parameters) Without Subscriptions
Key Questions Answered
Where the $1 and 18 cents come from?
The authors' motivation is down to earth. Greptile, a code review service, had charged one dollar for each pull request review since March. When agents on the team became more active, the monthly bill rose from $180 to $6451, roughly 36 times higher. Parcha then built its own reviewer on a cheaper model and tuned it on its own PRs. According to the authors, one review now costs 18 cents instead of a dollar. On 30 pull requests the system had never seen, it found 1.6 times as many real bugs as Greptile.
How it works?
AgentRun does not replace your agent. You choose the harness: Claude Code, Codex, OpenCode, or pi. The platform adds three things around it. The first is self-learning. After a run, the agent saves what will be useful next time: a note, a skill, or a ready-made playbook. Everything is stored in a separate git folder for the agent. Human feedback, such as a rating or a message after the first response, becomes lines in a corrections.md file, and every subsequent run receives it in full.
What you can use right now?
The platform is distributed under Apache-2.0. It requires Node 24.22.3 or newer, and runs on macOS or Linux. You can run it on a laptop, in a single container, or as a server with multiple runners. There are instructions for Railway, Render, a VPS, Kubernetes, Coolify, Dokploy, and Docker. If you already work with Claude Code or Codex, you can connect your existing configuration. For Codex and pi, signing in with ChatGPT is enough, with no model key needed.
Build & Ship Production AI Agents Without Boilerplate Code
Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.