Ellipsis vs Graphite (2026)

Upgraded from P. The June basis said the integration surface stayed concentrated on GitHub with no IDE or command line surface. Sentry adds error monitoring as a fourth class, the CLI adds a terminal surface, and per-agent MCP servers make the tool surface extensible by the customer rather than fixed by the vendor. Four event-source classes plus attachable MCP servers plus a browser IDE inside the sandbox meets breadth on the same standard applied to cubic.

Stands at F, re-based off first-party pages after the June basis cited git-tower.com. Recorded rather than credited to another axis: Cursor Cloud Agents now run inside the Graphite pull request, which is a post-acquisition integration and the clearest product evidence that the two are converging.

Upgraded from P. The June basis called the platform a fleet of independently triggered agents rather than one deep pipeline, which was fair then. What lifts it now is the agents-as-code lifecycle: the version on the default branch is the live agent, branches preview it, merging deploys with no deploy step, and git supplies review, history and rollback. Versioned, reviewable, revertible workflow definitions are the orchestration property the axis rewards, and it is a materially different design from warp's foreman or codebuff's role agents. Recorded honestly: coordination between concurrent agents is still not documented, so this is orchestration of definitions and sessions rather than of agents talking to each other.

Upgraded from P. The June basis called this human-collaborative and rules-based rather than an orchestration engine, sourced to a vendor blog post. Reading the plan matrix, the orchestration is real and layered: a stack-aware merge queue with basic and advanced tiers, an automations engine, a CI optimizer, and automatic rebasing of dependent branches as changes land. Sequencing dependent pull requests to keep trunk green while rebasing the chain is genuine multi-step workflow execution over a dependency graph, which is what the axis measures. Recorded honestly: this orchestrates the merge pipeline rather than coordinating multiple agents, so it earns F on the multi-step half rather than the multi-agent half.

Upgraded from P. The June basis described a GitHub-centred envelope with no IDE or command line surface; both halves have since changed. Sentry is a new fourth event source, an agent CLI now exists, and cron schedules are documented rather than inferred. Coverage now spans source control, chat, ticketing and error monitoring events, plus schedules, mentions, on-demand runs, an API and a CLI, which meets breadth on the same standard warp and goose were graded by.

Stands at P, re-based off the pricing matrix. Trigger coverage is real within one envelope: pull request events, merge queue events, CI events, plus CLI, VS Code, MCP, web and inbox surfaces, with Slack as a notification output. Held below F because everything is anchored to GitHub and the local development loop, with no chat-initiated or ticket-initiated invocation and no scheduled runs, which is the distinction separating this from warp, goose and ellipsis at F.

Stands at P, re-retrieved. The code search subagent combining keyword and vector search with tree-sitter AST parsing and pull request indexing is real and unusually well described for this lane, but it is engineering detail from the vendor's own build post rather than a grounding product, and no whole-repository index, cross-repo read or wiki surface is sold. That is the distinction that puts cubic and cosine at F and holds this at P: grounding here supports the workflow rather than being the product.

Stands at F but confidence is lowered from high to medium, because the June basis was graded at high confidence off braintrust.dev, a third-party vendor's customer case study, which the conventions exclude. Whole-codebase context and code indexing are confirmed first-party, but the mechanism is not documented anywhere I reached: no index architecture, retrieval method or scope statement appears on the pricing page or in navigation. The grade holds on the documented existence of code indexing plus admin controls over it; the confidence reflects that the depth claim rests on marketing rather than docs. A fetch of graphite.com/features/ai-reviews would settle it.

Upgraded from N. The June basis reasoned that learning which comments a team values was feedback adaptation belonging under Know, which was defensible, but the docs now document persistence mechanically: conversation state plus a workspace snapshot taken when a conversation goes idle, so the next message resumes with the working tree intact, and cached environment images carrying repositories and setup between sessions. Held at P rather than F because that is session resumption and environment reuse, not memory that accumulates or learns across sessions; the version on the default branch is the agent's whole persistent identity. Contrast cosine, where Memory is a named feature persisting architecture decisions, and warp, held at P because cross-harness memory is a research preview.

Stands at N on re-retrieval. The reviewer tuning itself from accepted and rejected comments is real but is feedback-loop adaptation of a model's behaviour rather than a documented memory layer, and imported style guides and rules are static configuration the customer maintains. Contrast cubic, moved to P because the docs name a Learns from you feature with a described mechanism, and cosine, moved to F for a named Memory feature persisting architecture decisions. Nothing comparable is named here. The distinction is the vendor documenting persistence as a capability rather than a training characteristic.

Upgraded from P, and the mechanism is structural rather than procedural, which is what makes it F. Most vendors in this lane gate agent actions by asking a human to approve them; Ellipsis narrows the credential so the disallowed action cannot be attempted. The vendor's own framing is that a narrowed token is minted by GitHub at the reduced scope, so nothing in the sandbox can exceed it, not a misbehaving tool and not a prompt injection hiding in a pull request. Budget caps enforced before a sandbox is created are the same shape. Pair with cubic's shadow-mode auto-approval and warp's approval gates on any comparison page: three genuinely different oversight designs.

Stands at P. Oversight is largely structural, since a reviewer's output is advisory findings a human accepts or rejects, which is the same reading applied to qodo. What lifts it toward the gate the axis wants is the merge queue: approved pull requests are sequenced and merged under rules the team configures, with advanced settings on Enterprise, and ACLs bound who can do what. Held below F because no per-action approval gate on the agent's own actions is documented, and no shadow or preview mode exists for the kind cubic ships on auto-approval.

Upgraded from P. The June basis conceded that SSO, SAML, SCIM, RBAC and audit were not enumerated; a dedicated security documentation page now enumerates the control surface in unusual depth. The axis conjunction is met: SOC 2 Type 1 is the attestation, and the named customer-facing controls are per-agent permission scoping down to read-only, write-only secret storage, org-admin configurable transcript retention from 7 to 365 days, and spend limits at four levels. Held at high confidence because this is documentation rather than marketing. Recorded rather than smoothed: Type 1 is a point-in-time design opinion rather than operating effectiveness, and no SSO or SAML is documented even now, so the identity half is still thinner than the isolation half.

Upgraded from N. The June basis said no enumerated identity and governance matrix is documented, sourced to the privacy and security docs page; the pricing page's Admin section enumerates it in full. The axis conjunction is met twice over: SOC 2 Type II plus continuous penetration testing on the attestation side, and SAML/SSO, ACLs, SIEM audit log export, code indexing controls, AI privacy controls and GHES support on the named-control side. This is the Sec understatement pattern the process note predicts, and it was found by reading the plan comparison table rather than the security page.

Upgraded from P and this is now among the strongest Obs cells in the lane, alongside warp and google-jules. The axis asks for the why and not just the what, and a retained per-step timeline with thinking tokens, tool calls, diffs and per-step cost is exactly that. Two details raise it above the usual: the record survives the sandbox being destroyed, and the vendor gives the buyer's own test for it, that which session produced a given pull request is a query months later. Replay is backed by retained model call records. Contrast augment-code, downgraded because streaming without retention is not a trace; here retention is the documented default.

Stands at P, re-based off the pricing matrix after the June basis cited a vendor blog post. Insights and custom analytics are genuine reporting, and the Enterprise SIEM audit log is a real governance artefact, but both measure people and access rather than agent behaviour. Under the lane-wide reading applied to augment-code and sourcegraph, reporting on throughput and cycle time is not a retained per-action trace of what the reviewer did and why. Would move to F on a documented per-review execution record.

Stands at P and the basis is now precise where it was vague. The private-instance-on-request claim came from the FAQ; what the security docs actually document is a hosted service whose model path can point at the customer's own AWS Bedrock or their own HTTPS proxy. That is control over where inference happens, not over where the platform runs, and the two are worth separating in the basis as they were for warp. No self-serve self-host, VPC install or region selection is documented, so this stays below F.

Stands at N, but the basis is corrected: the June note said no deployment option beyond hosted is documented, sourced to git-tower.com, and GitHub Enterprise Server support is in fact listed on Enterprise. That is not a deployment option for Graphite, though, it is support for a customer-hosted code host, so the grade does not move. This is the same distinction drawn on blink-new, where portability of the artifact did not confer residency on the platform. No region selection, VPC, private instance or self-host of Graphite itself is documented.

Stands at P, re-retrieved. Example agents are genuinely copyable YAML and the vendor names four of them, but they are starter templates in documentation rather than a catalogue: no marketplace, gallery or installable pack is documented, and there is no store to publish to. That is the same reading that held cubic at P for user-authored custom agents and lifted codebuff to F for a public Agent Store, so the three sit consistently.

Stands at N on re-retrieval, and this is a considered hold rather than a default. AI review customization covering automations, filters and rules is genuine reusable configuration, and it is what lifted comparable cells to P elsewhere; but the axis rewards prebuilt assets the customer installs or adopts, and every one of these is authored by the customer from scratch. Graphite ships a single reviewer agent with no library, gallery, template set or marketplace. Same reading that held cosine at N for vendor-built components. Searched: the pricing matrix, the features navigation and the homepage.

Upgraded from P. The June basis explicitly said Ellipsis does not support bring your own model API keys; a dedicated BYOK documentation page now exists, alongside customer-owned Bedrock and a custom LLM proxy. Four documented provider paths including the customer's own endpoint is the strongest form of this axis, matching how warp and qodo were graded. Recorded honestly and worth a comparison-page note: the managed default is Claude-only via Bedrock and the Anthropic API, and token-count metadata calls go to the Anthropic API on an Ellipsis key regardless of which provider serves completions, so the customer's choice is over the completion path rather than every call.

Stands at N on re-retrieval, and the temptation to move it was checked. AI privacy controls and code indexing controls are admin capabilities over data handling, not model selection, and the same distinction was applied to cubic, where bring-your-own API keys did not lift Model because supplying a key is not choosing a model. No provider is named as powering reviews on any first-party surface. Worth flagging as a likely mover: Cursor's stated plan on acquisition was to leverage its coding models to make Graphite's AI features more intelligent and to merge Graphite's reviewer with Cursor's Bugbot, so the model layer here may change materially.

Upgraded from P. The June basis said there was no documented public platform SDK, which was true of the product as it stood but is now contradicted on the docs landing page: a REST API at /v1, a TypeScript SDK on npm and a Python SDK on PyPI both generated from the API spec, and an agent CLI installed by Homebrew. The Agent Cloud launched 28 July 2026, after this record was built, so this is genuine product movement rather than a retrieval failure. MCP direction is inward here, servers the agent consumes, so the grade rests on the API, SDKs and CLI.

Upgraded from N. The June basis said no MCP server is documented, sourced to git-tower.com, a third-party blog; the pricing page lists MCP as a Hobby-tier feature on every plan, grouped under Stacking alongside the CLI and VS Code extension. Held at P rather than F because the grouping indicates an MCP server exposing Graphite's stacking operations to other agents rather than a general platform API, and no REST API, SDK or webhook surface is documented anywhere. F would need a documented API or SDK for building on Graphite.

DOWNGRADED from F, and this is the deliberate counterweight on a vendor where five cells moved up. The June F rested on dynamicbusiness.com, an excluded source, and credited bug detection plus sandbox-executed tested code as a testing capability. Under the lane-wide reading applied to google-jules, github-copilot and blink-new, an agent running the customer's tests and verifying its own output is P, not F; F requires a harness the customer can point at agent behaviour. Ellipsis ships no evals, benchmarking or regression surface. Analytics measures team velocity, which is throughput reporting rather than agent evaluation. The seventh application of this axis rule in this lane and the grade is now consistent with the other six.

Stands at F on the qodo and cubic precedent, where the axis and the product coincide: the customer points a shipped review product at their own code and it finds defects. Recorded honestly and worth noting against the seven downgrades this axis has taken in this lane: those were agents running a project's existing tests, whereas this is a defect-detection product with customer-configurable rules and filters. No harness for evaluating agent behaviour exists, which is the F bar on the other reading of this axis, so this F rests on the product-coincides reading alone.

Stands at P, and this is the one grid in the 30 June cohort where Comp was graded on the right basis rather than on sandbox execution, so it needed no correction. Re-retrieved and the evidence holds, with a browser IDE and previewable served ports now documented in the sandbox model. Held below F because the browser capability is a preview and inspection surface rather than a documented agent tool for operating third-party software without a programmatic interface, which is what cline and google-jules earned F for.

DOWNGRADED from P, the seventh correction of this identical axis error in the 30 June cohort. The June basis credited reviewing and navigating code, applying commits, rebasing branches via the CLI and self-healing CI as computer use; every one of those runs through git and the GitHub API, which are programmatic interfaces and exactly what the axis excludes. Both passes reached the pricing matrix, the features navigation and the homepage and found no browser, screenshot or GUI automation anywhere. Worth watching rather than assuming: Cursor Cloud Agents now run inside the pull request, and if those agents carry browser tooling this could change, but that capability belongs to cursor's record and is not documented on any Graphite surface.