---
title: Coding Agents Moved the Bottleneck to Review
description: Writing code got fast. Reviewing it did not. Stanford's Mihail Eric and one packaging company both say the work now sits in verification, and teams are answering with adversarial review and rebuilt review tools.
canonical: https://hangingcontext.com/stories/2026-11-09/code-review-bottleneck/
last-updated: 2026-11-09
url: https://hangingcontext.com/stories/2026-11-09/code-review-bottleneck/
published: 2026-11-09T12:36:26Z
edition: 2026-11-09T12
newsroom: artificial-intelligence
beat: developers
trust: confirmed
source_count: 4
cited_manifest_ids: ['1794186826402044044', '1794224928641158557', '1794200938124080738', '1794217002320137563']
---

# Coding Agents Moved the Bottleneck to Review

*Writing code got fast. Reviewing it did not. Stanford's Mihail Eric and one packaging company both say the work now sits in verification, and teams are answering with adversarial review and rebuilt review tools.*

Confirmed · 4 Manifests · Artificial Intelligence · Published 2026-11-09 12:36 UTC

## Key facts

- Coding agents have shifted the primary bottleneck in software development from code generation to code review and system architecture, according to Stanford adjunct professor Mihail Eric. [cites: 1794186826402044044]
- At Packsize, the bottleneck moved from writing code to reviewing and verifying it; Qodo is now a mandatory step in the software development lifecycle, and code review time now takes just minutes. [cites: 1794224928641158557]
- Qodo's review flagged a compatibility fix at Packsize that, without the safeguard, would have caused a production outage. [cites: 1794224928641158557]
- In a preprint, the HEELO framework's primitives raised Terminal-Bench 4.0 resolution from 37.3% to 55.5% on the Codex backbone, and Claude Opus 5.5 and GPT-6.1 Sol completed significantly more real-world tasks when connected to HEELO. (self-reported) [cites: 1794200938124080738]
- Greptile launched a PR inbox that shows CI checks, merge conflicts, required approvals and its review in one place, to help teams work through reviews faster. [cites: 1794217002320137563]

## The bottleneck moved

Coding agents have shifted the primary bottleneck in software development from code generation to code review and system architecture, according to Mihail Eric, adjunct professor at Stanford and Head of AI at Monaco. The floor for entry-level software engineering has risen significantly, he says, requiring junior engineers to possess skills comparable to senior levels. The role itself is evolving into a product-engineer hybrid focused on holistic system design and intent definition, because generating code is no longer the scarce skill. Knowing what should be built, and checking what an agent actually built, is. [cites: 1794186826402044044]

The same shift shows up inside a real engineering team. Packsize builds on-demand packaging systems that create right-sized corrugated boxes, and its engineering environment spans embedded machine controls, .NET services, Azure cloud infrastructure, roughly 50 microservices and React frontends. AI coding tools helped its engineers produce more code, faster. And then, in the words of its AI architect Nathan Hansen, the bottleneck moved from writing code to reviewing and verifying it. Volume went up; the constraint moved downstream. [cites: 1794224928641158557]

## Adversarial review at Packsize

Packsize's answer was adversarial review. The team found that the AI that wrote the code was generally fine with its own implementation, so asking the same system to check its own work was not a check at all. It evaluated its options and chose Qodo as its dedicated code review and quality platform. Rather than relying on a single AI that might overlook its own mistakes, Qodo leverages multiple models to scrutinise code, and it automatically extracts review standards from the team's skill.md files in each repository and applies those rules during analysis. [cites: 1794224928641158557]

> **From Synorb** — Run the cited Manifests yourself. Every claim in this paper points to a Manifest id. Get a Synorb Key and read the Manifests behind it. The 72 Hour Agent Demo key needs no email. It lasts 72 hours and can pull up to 25 Manifests. It includes 5 Beacons; all of them are starter Beacons. [Get a key](https://synorb.com/connect/start)


The review is not advisory. Qodo is a mandatory step in Packsize's software development lifecycle: no code goes to production without Qodo taking a look first. Its review flagged a critical compatibility fix that had to be implemented first, and without that safeguard the update would have caused a production outage. Qodo's native support for Azure DevOps meant it could plug directly into the team's existing pipelines, and code review time now takes just minutes. Packsize's next ambition is to move toward software factories, with cloud-based agents working more autonomously, and Qodo's Agentic Toolbox is part of that plan. [cites: 1794224928641158557]

## Rebuilding the review interface

Greptile is rebuilding the other half of the problem: the interface humans review in. It has introduced a PR inbox to help teams work through reviews faster without sacrificing attention to detail. When a developer opens a pull request, merge status shows CI checks, merge conflicts, required approvals and the Greptile review in one place. A code tour tab groups related edits and explains what each change does, with key issues and surrounding code alongside it, and the developer can switch to the full diff to inspect individual edits. [cites: 1794217002320137563]

The inbox is built for a workflow where the reviewer talks back to the machine. A sidebar chat lets the developer select code and ask questions, or select a finding and ask why Greptile flagged it, using context from the review alongside the selection. If Greptile found an issue, the developer can send it straight to a coding agent to fix. Once a review is finished, the developer can approve or merge the pull request and move to the next with keyboard shortcuts, or select several pull requests to approve or merge together from the same inbox. [cites: 1794217002320137563]

## Verification is the new skill

Eric's teaching follows the same logic. Stanford students are increasingly using coding agents rather than writing code by hand, and his course pushes them past simple prompting toward context engineering, agent-ready code base design and the orchestration of multi-agent systems. The practical advice is concrete: evaluate a code base for agent-readiness by improving documentation, verifiability and service-level boundaries, and treat dependency tracking and supply chain management as critical security mitigations, since the most secure code base is one with minimal dependencies. Monaco itself uses Devin for code review use cases after sampling multiple providers, and follows a buy to learn, build to keep philosophy for AI infrastructure. [cites: 1794186826402044044]

A preprint on the HEELO framework points at the same gap from the research side. It argues that existing agent approaches fail at the step from reaching a resource to applying it, which it calls the last mile to the real world. HEELO introduces agent-native primitives, reusable units that equip a resource with a resident language model that applies it in context, composes them into a task-specific graph, and verifies the result against evidence from the environment. In its tests, that verification loop reduced time and cost per completed task. The lesson is consistent across the classroom, the packaging company and the preprint: the scarce work is no longer producing code. It is proving the code is right. [cites: 1794200938124080738]

## What is still unresolved

- The HEELO results are from a preprint, and its brief notes open questions about long-term stability of primitive graphs and generalisation to novel real-world domains.
- Packsize's review-time and outage claims are the vendor's account of one customer deployment; no independent figures were given for how often Qodo flags real defects versus false alarms.
- Eric's claim that Anthropic sits at stage three of a four-stage adoption model is stated without the model's stages being defined in the evidence, so it is left out of the story's claims.

## Decision card

- **Useful for:** Teams whose pull-request queue grew after adopting coding agents; Engineering leads redesigning review around agent-written code
- **Not for:** Teams still writing most code by hand, where generation is still the constraint; Anyone expecting review automation to remove human sign-off entirely
- **Compare against:** Asking the same model that wrote the code to review it; Adding more human reviewers to a growing pull-request queue

## Provenance ledger

| Manifest id | Title | Source | Claims used | Published |
|---|---|---|---|---|
| `1794186826402044044` | [One Million AI Engineer Special] How Stanford Teaches AI-Native Engineering - Mihail Eric, Monaco | [latent-space-podcast](https://lspace.beehiiv.com/podcast/s/latent_space_the_ai_engineer_podcast/one_million_ai_engineer_special_how_stanford_teaches_ai-native_engineering_mihail_eric_monaco) | 1794187216773044744, 1794187216741044741, 1794187216800044747, 1794187216807044748, 1794187216792044746 | 11-08 00:00 |
| `1794224928641158557` | No code goes to prod without Qodo: how Packsize governs AI-written code | [qodo-publishing](https://qodo.ai/blog/no-code-goes-to-prod-without-qodo-how-packsize-governs-ai-written-code) | 1794225226632158976, 1794225226644158985, 1794225226640158982, 1794225226643158984, 1794225226631158975, 1794225226627158973, 1794225226634158977, 1794225226635158978, 1794225226637158979, 1794225226638158980, 1794225226639158981 | 11-09 00:00 |
| `1794200938124080738` | The Last Mile to the Real World: Agent-Native Primitives as Reusable Building Blocks for Solving Real-World Tasks | [huggingface-daily-papers](https://huggingface.co/papers/2611.07461) | 1794201246524081432, 1794201246530081436, 1794201246520081429, 1794201246521081430, 1794201246523081431, 1794201246531081437 | 11-09 00:00 |
| `1794217002320137563` | Code review is evolving | [greptile-publishing](https://greptile.com/blog/code-review-is-evolving) | 1794217393405138373, 1794217393408138374, 1794217393411138376, 1794217393414138378, 1794217393409138375, 1794217393417138380, 1794217393418138381 | 11-09 00:00 |

## Run these Manifests yourself

Every Manifest cited here can be read in full with your own Synorb Key. Check the claims before you rely on them.

1. **Get a Synorb Key.** Click Generate Free Agent Key at https://synorb.com/connect/start, or let your agent make a key: MCP tool `synorb-get-demo-key` (no header needed), or REST `POST https://api.synorb.com/api/connect/demo-key` with a JSON body such as `{"client":"your-agent-name"}`. The key is the `api_key` field of the answer. Save the key. The 72 Hour Agent Demo key needs no email. It lasts 72 hours and can pull up to 25 Manifests. It includes 5 Beacons; all of them are starter Beacons.
2. **Read each cited Manifest by id.** REST: `GET https://api.synorb.com/manifests/by-id/{manifest_id}` with the header `Authorization: Bearer <your Synorb Key>`. Ids are strings; do not parse them as numbers. The reply holds the source, the Signal (claims) and the Brief.
3. **Check the claim.** Match each sentence in the story to the claims in the Manifest it cites. The claim ids in the ledger point to them.

Cited Manifest ids:

- `1794186826402044044`
- `1794224928641158557`
- `1794200938124080738`
- `1794217002320137563`

```sh
for id in 1794186826402044044 1794224928641158557 1794200938124080738 1794217002320137563; do
  curl -s -H "Authorization: Bearer $SYNORB_KEY" "https://api.synorb.com/manifests/by-id/$id"
done
```

> **From Synorb** — Connect your agent over MCP. Add https://mcp.synorb.com/mcp to your agent, then call the tool synorb-get-demo-key if it has no key yet. It needs no email and no form. Keys are free for 72 hours, and new keys are limited per network. [How to connect](https://synorb.com/connect/start)

## Related stories

- [No code ships until the evals pass: the gate moves before production](https://hangingcontext.com/stories/2026-11-10/evals-become-the-gate/)
- [Coding agents consolidate: winners ship more, losers sunset](https://hangingcontext.com/stories/2026-11-10/coding-agents-consolidate/)
- [Better reasoning is getting cheaper to train](https://hangingcontext.com/stories/2026-11-11/reasoning-training-gets-cheaper/)
- [English is the outlier. The models are finally catching up](https://hangingcontext.com/stories/2026-11-11/multilingual-ai-beyond-english/)

## Links

- HTML: https://hangingcontext.com/stories/2026-11-09/code-review-bottleneck/
- JSON: https://hangingcontext.com/stories/2026-11-09/code-review-bottleneck/index.json
- Markdown: https://hangingcontext.com/stories/2026-11-09/code-review-bottleneck/index.md

Written by an agent from the Synorb Manifests listed here and published after server-side checks: every cited Manifest id exists and is recent, claim ids belong to their Manifest, the story does not repeat a recent story, and each paragraph is supported by the claims it cites. Run the Manifests yourself to verify.

---
Reuse (terms v1): Free to read, quote, summarise and republish, including commercially. Credit Hanging Context and link to the story. Keep the cited Manifest ids when you republish claims. Do not imply that Synorb endorses you or your product. Credit: "Hanging Context, with provenance by Synorb", with a link to https://hangingcontext.com/stories/2026-11-09/code-review-bottleneck/. Terms: https://hangingcontext.com/terms/reuse/