Agent Metrics
Which AI coding agent is fastest?
Turn times and usage for 15 AI coding agents, from 6.5M+ turns in Zed. Updated weekly.
Last updated: October 5th, 2026Key Findings
- GitHub Copilot had the fastest median turn time among the 10 most-used agents, at 39.6 seconds.
- External agents accounted for 65% of sessions, led by Claude Agent (268K) and Codex CLI (114K).
- Local models ranked second among models used with Zed Agent, with 24K sessions.
Fastest AI Coding Agents by Turn Time
Median (p50) turn time for the selected agents over the last 30 days, fastest first. p10 and p90 show the spread.
Explore Full MethodologyMost Popular AI Coding Agents
Here's how weekly are trending across the selected agents, over the last 30 days.
Explore Full MethodologyTotal Sessions
779.5KTotal Turns
6.5M| 1 | 277,257 | ||
| 2 | 268,046 | ||
| 3 | 114,436 | ||
| 4 | 58,877 | ||
| 5 | 18,046 | ||
| 6 | 16,783 | ||
| 7 | 12,636 | ||
| 8 | 8,265 | ||
| 9 | 3,006 | ||
| 10 | 2,167 |
Local Models in the Mix
Local models run entirely on your device, as opposed to cloud-hosted ones.
Learn MoreLocal Sessions
13.6KLocal Turns
61.4KCross-ecosystem by design
Developers can bring any AI agent they want to Zed (we think of ourselves as the "Switzerland" of editors). Many agents connect through ACP, the Agent Client Protocol, a shared open standard for agent-editor communication.
This means we see agent usage across the ecosystem, not just our own. We can offer a more comprehensive view of how AI agents are actually used than most editors can — or would.
Frequently Asked Questions
Anonymized and aggregated data from users interacting with the Zed agent and other agents via the Agent Client Protocol. No individual user data is exposed.
We re-pull data weekly. This allows time for new models to accumulate meaningful sample sizes.
We only show agents that have sufficient data to report meaningful metrics. Agents with very low usage within Zed may not appear.
No. This is aggregated, anonymized data. We do not track or expose individual user metrics.
Turn time is how long an agent takes to finish responding after you send it a message. p50 is the median: half of all turns finish faster than this. p10 is the time the fastest 10% of turns finish within, and p90 is the time 90% of turns finish within. The gap between p50 and p90 shows how long an agent's slowest tasks run.
Turn time depends on model configuration, context window size, task complexity, and infrastructure factors. These metrics show aggregate trends, not guarantees.
Methodology
The metrics on this page are derived from anonymized, aggregated telemetry collected from Zed users who interact with AI agents. No individual session data, user identifiers, or proprietary code is exposed, and all data is aggregated before display.
-
What we collect: When you use an AI agent in Zed, we record metadata about the interaction: which agent was used, how long it took to respond, whether you accepted or rejected the suggested edits, and how many lines of code were involved. For Zed's own agent, we also capture which underlying model powered the response. We do not collect the content of your prompts, the code you're working on, or any personally identifiable information.
-
What we exclude: To ensure the data reflects real-world usage, we exclude all interactions from Zed staff accounts and from Zed Nightly builds (which may contain experimental or unstable behavior). We also apply minimum thresholds: agents and models with very low usage volumes are not displayed to avoid misleading statistics from small sample sizes.
-
How we calculate metrics: Turn time is measured from request initiation to response completion in milliseconds. Error rate is the percentage of turns that resulted in a failure status.
-
When we update: Data is refreshed every week. This cadence allows newly released models and agents to accumulate enough usage for statistically meaningful comparisons while keeping the data reasonably current.
-
Known limitations: Model-level breakdowns are only available for Zed's own agent. External agents like Claude Agent and Codex don't reliably expose which model they're using in their telemetry. We also cannot detect user subscription tiers (e.g., whether someone is using Gemini's free or paid tier), which may affect performance characteristics.