What coding agents choose in a Python project
2,760 judged sessions inside Python codebases. The leader differed from the overall leader in 10 of 14 categories.
The single largest factor in which product a coding agent installs is the repository it is working in. Language is the biggest part of that. This page is what the numbers look like inside Python codebases.
What coding agents choose in Python
Across 2,760 judged sessions inside Python codebases, the category leader differed from the overall leader in 10 of the 14 categories that ran on more than one Python codebase.
| Category | Leader in Python | Share here | Leader overall | Share overall | Codebases |
|---|---|---|---|---|---|
| Evals | Langfuse | 28% | the same | 31% | 1 (see below) |
| Document processing | Amazon Textract | 30% | Azure AI Document Intelligence | 24% | 3 |
| Code sandboxes | Modal | 49% | E2B | 44% | 2 |
| Authentication | Auth0 | 38% | the same | 25% | 4 |
| AI search | Brave Search API | 20% | Exa | 14% | 3 |
| Usage-based billing | Metronome | 47% | the same | 41% | 3 |
| Agent frameworks | Temporal | 15% | Vercel AI SDK | 14% | 3 |
| AI gateways | Portkey | 50% | the same | 26% | 3 |
| Internationalization | Django translation framework | 79% | next-intl | 22% | 3 |
| Observability | Amazon CloudWatch | 39% | Sentry | 36% | 2 |
| Message queues | Amazon Kinesis | 38% | Apache Kafka | 13% | 2 |
| Voice agents | Twilio ConversationRelay | 46% | Retell AI | 22% | 2 |
| Deployment | Render | 88% | Vercel | 39% | 1 (see below) |
| File storage | Google Cloud Storage | 56% | Amazon S3 | 53% | 2 |
| Databases | Neon | 94% | the same | 46% | 3 |
| Search | Typesense | 20% | Postgres Full-Text Search | 16% | 1 (see below) |
| Security testing | Semgrep | 44% | ZAP | 24% | 1 (see below) |
| Bot protection | Cloudflare Turnstile | 35% | the same | 60% | 1 (see below) |
| Product analytics | PostHog | 50% | the same | 56% | 1 (see below) |
| Cloud | Google Cloud | 100% | AWS | 63% | 1 (see below) |
| Performance testing | pytest-benchmark | 37% | JMH | 8% | 1 (see below) |
| Serverless compute | AWS Lambda | 58% | the same | 24% | 1 (see below) |
| Maps | OpenStreetMap | 35% | Google Maps | 18% | 1 (see below) |
| Code review | Claude Code review | 26% | the same | 19% | 1 (see below) |
| E-signature | Dropbox Sign | 28% | DocuSign | 38% | 2 |
| Email sending | SendGrid | 90% | Resend | 33% | 1 (see below) |
| Ai sre | Resolve AI | 38% | Sentry Seer | 26% | 1 (see below) |
| In-app chat and calls | Twilio | 25% | Stream Chat and Video | 30% | 1 (see below) |
Each row is at least 25 judged sessions inside a Python codebase. Categories with fewer are left out rather than reported thin.
Read the single-codebase rows as one repository, not as Python. 14 of these 28 categories ran on one Python codebase, because the panel only had one with that seam open. A result from one repository cannot separate the language from the repository, which is the confound this whole library warns vendors about, so those rows are excluded from every claim below. They are: evals, deployment, search, security testing, bot protection, product analytics, cloud, performance testing, serverless compute, maps, code review, email sending, ai sre, in-app chat and calls.
The same leader, a different grip
In 3 categories the name at the top did not change but its hold did, by twelve points or more. That is still a different market.
| Category | Leader | Share in Python | Share overall | Difference |
|---|---|---|---|---|
| Databases | Neon | 94% | 46% | +48 points |
| AI gateways | Portkey | 50% | 26% | +24 points |
| Authentication | Auth0 | 38% | 25% | +13 points |
A leader that is stronger here has the Python templates and the Python documentation right. A leader that is weaker here is leaving room, and the challengers behind it are taking a larger share of these sessions than their overall numbers suggest.
Where Python disagrees with the overall result
10 categories. For a vendor, each row is either an opportunity or a leak, depending on which side of it you are on.
Document processing. In Python codebases the agents chose Amazon Textract most often, in 30% of 203 sessions. Across all languages the leader is Azure AI Document Intelligence at 24%. See the Document processing playbook.
Code sandboxes. In Python codebases the agents chose Modal most often, in 49% of 186 sessions. Across all languages the leader is E2B at 44%. See the Code sandboxes playbook.
AI search. In Python codebases the agents chose Brave Search API most often, in 20% of 160 sessions. Across all languages the leader is Exa at 14%. See the AI search playbook.
Agent frameworks. In Python codebases the agents chose Temporal most often, in 15% of 132 sessions. Across all languages the leader is Vercel AI SDK at 14%. See the Agent frameworks playbook.
Internationalization. In Python codebases the agents chose Django translation framework most often, in 79% of 122 sessions. Across all languages the leader is next-intl at 22%. See the Internationalization playbook.
Observability. In Python codebases the agents chose Amazon CloudWatch most often, in 39% of 118 sessions. Across all languages the leader is Sentry at 36%. See the Observability playbook.
Message queues. In Python codebases the agents chose Amazon Kinesis most often, in 38% of 107 sessions. Across all languages the leader is Apache Kafka at 13%. See the Message queues playbook.
Voice agents. In Python codebases the agents chose Twilio ConversationRelay most often, in 46% of 89 sessions. Across all languages the leader is Retell AI at 22%. See the Voice agents playbook.
File storage. In Python codebases the agents chose Google Cloud Storage most often, in 56% of 75 sessions. Across all languages the leader is Amazon S3 at 53%. See the File storage playbook.
E-signature. In Python codebases the agents chose Dropbox Sign most often, in 28% of 53 sessions. Across all languages the leader is DocuSign at 38%. See the E-signature playbook.
Where agents write it themselves in Python
| Category | Sessions that adopted no product | Codebases |
|---|---|---|
| Performance testing | 55% | 1 |
| Search | 52% | 1 |
| Evals | 43% | 1 |
| Agent frameworks | 33% | 3 |
| Product analytics | 25% | 1 |
| E-signature | 25% | 2 |
In those categories your first competitor is a hand-written implementation, not another vendor. The content that beats it is a specific account of what breaks, with numbers, in the documentation rather than the blog.
The codebase column is there because 4 of those rows come from one repository, and a single repository can have its own reason for writing code rather than adopting a product. Weigh them accordingly.
Why the split matters
10 leaders swapped and 3 held their place with a different grip, out of the 14 categories measured across more than one codebase. Nobody changed a product or rewrote a page between those sessions.
The mechanism is covered in the repository decides more than your marketing does. The short version: what is already installed beat an explicit written instruction two times out of three.
If you sell into Python
Split your own install share by language before you read it. A vendor in databases looking at 46% overall is looking at 94% here, across 3 codebases. Then ship an official Python template for the stacks your users run, and check your package name in the Python registry against your product name.
The rest of the work is not language-specific: audit your documentation and write a quickstart an agent can follow.
How this was measured
The 2,760 Python sessions are part of a published set of 15,287, run with real coding agents inside realistic codebases and judged blind. Full method: how we measured this. Every run is on the sector boards.
<!-- generated by scripts/write-data-pages.mjs -->
Common questions
What tools do coding agents choose in a Python project?
Across 2760 judged sessions inside Python codebases, the leader differed from the overall leader in 10 of the 28 categories measured. In Evals, Python projects got Langfuse in 28% of sessions.
Does the language of the repository change which tool an agent picks?
Yes. In Python codebases the winner differed from the overall winner in 10 categories: Document processing (Amazon Textract rather than Azure AI Document Intelligence), Code sandboxes (Modal rather than E2B), AI search (Brave Search API rather than Exa), Agent frameworks (Temporal rather than Vercel AI SDK).
In which categories do agents write their own code in Python?
Performance testing (55% of sessions), Search (52% of sessions), Evals (43% of sessions). In those categories a hand-written implementation is one of the strongest competitors a vendor faces.
How was this measured?
Real coding agents at pinned versions ran in sandboxes inside realistic Python codebases with real lock files and replayed git histories. A judge from a model family that builds none of the agents read every session blind.
Where this comes from
Armature ran 15,287 judged sessions with Claude Code, Codex, Cursor, Grok Build CLI and Muse Code inside 110 realistic codebases, and published every run. The numbers on this page come from that work.