Agentic Index
Best coding agents ranked on the merge loop (2026)
The Agentic Index grades 63 coding agent vendors against the same 14 capabilities as every other lane. These are independent ratings of AI coding agents, rated from each vendor's own public evidence rather than from analyst opinion, vendor briefings or paid placement. 4 of them document the full merge loop. That means the agent reads the codebase, reaches the systems the work actually lives in, verifies its own change before a human sees it, leaves a record somebody can read afterwards, and a human governs what reaches the default branch. That is 6.3% of the lane, the strictest result this index has produced. This is a bar, not a leaderboard: a vendor either documents all five in its own public materials or it does not clear, and Partial evidence on any one does not count.
Before the list, the thing worth knowing about the list. None of the 4 vendors that clear this bar writes the code. Baz is code review, Sonar is code verification, Snyk and depthfirst are security. The entire cleared set is the layer that checks the work. Every vendor whose product is writing code fails, including all six of the highest coverage vendors in the pool: Anthropic Claude Code 13.5, Google Antigravity 13.0, OpenAI Codex 13.0, Cline 12.5, GitHub Copilot 12.5, OpenHands 12.5, every one of them above the 12.0 carried by the top of the cleared list. Four of those six fail on the same single capability. The category's leaders are broad and unverified, and the vendors that clear a verification bar are the vendors whose product is verification.
The finding underneath that is the reason the third capability is on the bar. This field can write the code and cannot prove it works. Testing, debugging and optimization sits at 18 of 63, the thinnest axis on the bar by a distance, and it is the sole blocker for 10 of the 18 vendors sitting one capability short. 37 vendors document both reading the codebase and reaching your systems in full, and 25 of those 37 do not document self verification. That is the cleanest single test to run in a demo: ask to watch the agent run the suite, read a failure and fix its own work with nobody prompting it.
Two more numbers to carry into a vendor call. The field asks permission and then does not check its work. 33 of 63 document human oversight and guardrails in full, nearly twice the 18 documenting self verification, and 23 of those 33 document no testing. That ordering is backwards for the job. An approval gate in front of an unverified change does not reduce review burden, it moves the whole verification task onto the reviewer, who is the person the agent was bought to spare. And memory is finally present. 26 of 63 document memory and state persistence in full, 2 of the 4 that clear, which breaks a run of three consecutive lanes where persistent state was nearly absent. That is what you would expect from a category built on repository context, and it is worth naming as the exception rather than the rule.
One honest note on the bar itself. 6.3% is harsh, and it is deliberate. Drop self verification and the list gets much longer and much less useful, because a coding agent that writes a change and cannot document checking it has handed a reviewer the exact part of the job they were delegating. The bar stays at five for that reason. Disagree with it if you like: every count on this page is stated so you can recompute the result against four.
The merge loop, and how the 63 vendors score against it
| Step | Capability, and what has to be documented | Full | Only blocker |
|---|---|---|---|
| 1.It reads the codebase | Knowledge grounding and RAGThe agent grounds itself in the actual repository, its conventions and its history, rather than reasoning over the file currently open. Without this it is guessing in your dialect. | 40 (63%) | 0 |
| 2.It reaches the systems | Integrations and tool callingIt calls into the repository, the build, the test runner, the issue tracker and CI, so the work lands where the work lives rather than in a chat window. | 54 (86%) | 1 |
| 3.It verifies its own change | Testing, debugging and optimizationIt runs tests, reads the failure, retries and gates its own output before a human sees it. This is the capability that separates an agent from an autocomplete, and it is the thinnest on this bar. | 18 (29%) | 10 |
| 4.It leaves a record | Observability and auditabilityA durable trace of what the agent read, what it changed and why, readable by the person reviewing the diff and by whoever inherits the code a year later. | 23 (37%) | 5 |
| 5.A human governs the merge | Human oversight and guardrailsApproval gates, scoped permissions or branch protection on anything that reaches the default branch. Shipping to main is not a suggestion. | 33 (52%) | 2 |
Full means the vendor publishes evidence meeting the capability in its own public materials, under the Agentic Index verification standard. Only blocker counts vendors that document the other four capabilities in full and fail on this one alone. Read the last two columns in order: tool calling reaches 54 of 63, grounding 40, oversight 33, then auditability drops to 23 and self verification to 18, where it accounts for 10 of the 18 near misses on its own. The field can read your repository and change it. What it cannot consistently document is checking the change.
The 4 vendors that document the full merge loop
Ordered by total coverage across all 14 capabilities, ties broken alphabetically, which is exactly what the rankings pages compute. Coverage measures breadth across the whole taxonomy, so a lower score here usually means a narrower product rather than a weaker one. Every entry says what it actually operates on, beside its name, because all four run this loop over somebody else's code rather than writing their own.
-
1.Baz
12.0 / 14 capabilities
Operates on code by: reviews and governs code before and after it is written, including code written by other agents
The highest coverage score on this list and the only cleared entry whose primary category is Coding agent rather than security. Baz runs agents directly on the codebase to review, govern and secure what developers and AI copilots produce: a code review agent on every pull request that detects breaking changes across endpoints and behaviour, a Spec Reviewer that runs the app in a sandbox and drives a browser to check the built experience against Figma and Jira, and Planner, which reviews and approves the plan before any code is written. The founding team came out of Palo Alto Networks by way of Bridgecrew, and it extended a seed round to seventeen million dollars in June 2026 co led by Battery Ventures and Boldstart. Its stated premise is the one this page keeps arriving at from the data: when AI generates code faster than a team can read it, reviewing after the fact stops being a defence. Pricing is a free trial and an on demand starter on the GitHub Marketplace, with team and enterprise tiers not fully published.
-
2.Snyk
11.5 / 14 capabilities
Operates on code by: secures and governs autonomous coding agents operating inside your codebase (secondary lane membership, primary category Security / SOC agent)
In this lane by secondary category, and the clearest statement of what the cleared set actually is. Snyk's Evo agents exist to govern other vendors' coding agents: AI-SPM visibility across models, agents, MCP servers, skills and tools with an AI bill of materials showing what each one can reach, continuous autonomous pentesting, and runtime enforcement through Agent Guard. Its Agent Security solution names Claude Code, Cursor and Devin as the things being governed, which is a fair summary of the finding on this page. More than 4,800 customers. Pricing is published in full and starts free, with Team at twenty five dollars per contributing developer per month, which in a lane this quiet on pricing is close to a differentiator on its own.
-
3.Sonar
11.5 / 14 capabilities
Operates on code by: verifies code against a deterministic rule set, whether a human or an agent wrote it
The incumbent verification layer, and it grades like one. SonarQube is used by more than seven million developers and over twenty two thousand customers including most of the Fortune 100, and its pitch for the agentic era is a neutral trust layer holding all code to one standard regardless of who wrote it. The distinction that matters against the newer LLM reviewers is that its core analysis is deterministic and rule based, so findings are transparent, auditable and defensible in a compliance conversation rather than a second opinion from a second model. Its open source MCP server lets a coding agent verify its own output mid loop, which is the merge loop running inside the agent rather than after it. It acquired Gitar in May 2026 to add narrative review on top of the engine. Free tier and a 14 day trial, paid plans priced by lines of code.
-
4.depthfirst
11.0 / 14 capabilities
Operates on code by: detects, triages and remediates vulnerabilities across code, dependencies and runtime (secondary lane membership, primary category Security / SOC agent)
The newest and the narrowest, and in this lane by secondary category. depthfirst is an applied AI lab building security specific models rather than bending general ones to the task, with agents that detect, triage and remediate vulnerabilities across code, dependencies, secrets, infrastructure and runtime, plus agentic penetration testing. Founded in 2024 by Databricks' former head of infrastructure security and a seven year DeepMind researcher, it has raised 120 million dollars, including an 80 million dollar Series B led by Meritech in March 2026 at a 580 million dollar valuation, less than ninety days after its Series A. Pricing is not published and it sells through enterprise sales scoped to codebase size and developer count, so size it early rather than late.
The 18 vendors that miss by exactly one capability
These document four of the five and fail one. In this lane it is also the list of every name you have heard of, and Anthropic Claude Code sits at the top of it carrying the joint highest coverage score in the entire pool. If you are already holding a shortlist this table is the most useful thing on the page, because it tells you the single question to ask each vendor rather than sending you back to the start. A miss is a documentation finding, not a verdict on the product: several of these vendors may well do the thing and have not published evidence that meets the standard.
Common questions
What are the best coding agents in 2026?
Four of the 63 vendors in the Agentic Index coding agent lane document the complete merge loop in full: the agent reads the codebase, reaches the systems the work lives in, verifies its own change, leaves a readable record, and a human governs the merge. Ordered by total coverage across all 14 capabilities they are Baz, Snyk, Sonar and depthfirst. That is 6.3% of the lane, the strictest result this index has produced. Note what those four have in common: none of them writes the code. They review, verify and secure it. Eighteen more vendors sit exactly one capability short, and that table is where every name you have heard of appears.
Why does Claude Code, Cursor or GitHub Copilot not appear on the list?
All three document four of the five capabilities and fail the same one, which is testing, debugging and optimization. So do OpenAI Codex, Google Jules and Poolside. Anthropic Claude Code carries the joint highest total coverage in the pool at 13.5 of 14, above every vendor that clears the bar, and still does not clear it. That is not a claim that these tools cannot run a test. It is a finding about what each vendor publishes: the evidence for grounding, tool calling, oversight and auditability is on the record, and the evidence for the agent gating its own output before a human sees it is not. If you use one of these, the question to put to the vendor is narrow and answerable.
Why is self verification on the bar at all?
Because it is the capability that separates an agent from an autocomplete. A coding agent that writes a change and cannot show that it checked the change has handed the reviewer the exact part of the job the reviewer was trying to delegate. It is also the thinnest axis in the lane by a distance, documented in full by 18 of 63, and the sole blocker for 10 of the 18 vendors that otherwise clear. Running the same pool against a four capability bar without it produces a much longer and much less useful list, which is why the bar stays at five and why this answer exists.
How many coding agents can actually act on a repository?
Thirty seven of the 63 document both knowledge grounding and integrations and tool calling in full, meaning they read the real codebase and reach the repository, build, tests and tracker. Of those 37, only 12 also document self verification. So 25 vendors can make a change in your systems and do not document checking it first. That is the single cleanest question to take into a demo: ask to watch the agent run the test suite, read a failure and fix its own work without anybody prompting it.
Do these tools have human approval before code is merged?
Thirty three of the 63 document human oversight and guardrails in full, which is nearly twice the number documenting self verification. That ordering is backwards for the job. An approval gate in front of an unverified change does not reduce review burden, it relocates the whole verification task onto the reviewer. Twenty three of the 33 vendors that document oversight do not document testing, and those 23 are where the review load is quietly landing.
How is this different from a coding agent benchmark or leaderboard?
A benchmark measures how well a model completes tasks. This measures what a vendor documents about how its agent operates inside your engineering process, graded against a fixed 14 capability taxonomy from the vendor's own public materials, with no vendor briefings, no submissions, no paid placement and no vendor review before publication. The bar is mechanical and stated in full so you can disagree with it and recompute it. It also names what each vendor fails rather than only what it does, which is why the near miss table is four times longer than the list above it.
Method. This page runs the horizontal grading method over one lane of the index: membership is the Coding agent category, 63 vendors out of 976 researched, 48 by primary category and 15 by secondary, every one carrying a complete 14 row capability grid, with none excluded for incomplete evidence. The bar is five of those 14 capabilities at Full, chosen because they are the loop a change actually travels, and it is stated in full above so you can disagree with it and recompute. Every grade comes from the vendor's own public materials under the Agentic Index verification standard. No vendor pays for placement and no vendor has reviewed this page. Data last verified September 7, 2026. How this evidence is graded
Related: all 63 coding agents ranked on total coverage, the full coding agent capability matrix, how every vendor scores on testing and self verification, how every vendor scores on knowledge grounding, Cognition against Factory, Charm against OpenCode, CodeRabbit against Graphite, Cognition against Cursor, agent observability platforms, where the record gets read, platforms that require human approval before an agent acts, production ops agents ranked for automated incident resolution, AI SOC agents for alert triage, compare vendors side by side.