Cosine vs Poolside (2026)

Breadth across classes is met on the named set alone: source control, ticketing, database, design, payments and chat are six distinct classes.

Centrally managed, admin-approved MCP servers reach Slack, Jira, GitHub, Workday, Salesforce, Linear and Notion, and agents edit files, run commands and push to source control inside governed sandboxes.

Swarm mode is a named orchestrator spawning specialized child agents that work in parallel, which is coordinated multi-agent work rather than merely parallel runs.

The Console builds single and multi-agent pipelines in sandboxes, with a worked example running from specification to a passing build, and Poolside acquired Fern Labs for its Bridge orchestration layer.

Work starts from three surfaces, the terminal CLI, Cosine Cloud and Desktop, each invoked by a person, with remote execution behind them. GitHub, Jira, Linear and Slack appear as MCP connections the agent calls rather than as channels that invoke it, and no IDE extension is documented. A documented ticket, webhook or scheduled trigger would move this to Full.

GitHub Actions runs Poolside on pull request events, a nightly cron schedule or on demand, alongside the CLI, editors, the desktop Assistant and the API.

Context is gathered on demand: Cosine loads relevant files and uses language server operations such as go to definition and find references, with MCP connections for outside systems. No persistent index or embeddings layer over the codebase is documented, which keeps this at Partial.

Context is assembled per run from the working directory, local skills, MCP servers and web search; no maintained retrieval structure over the customer's knowledge is documented.

The agent saves reusable facts with a save_memory tool into .cosine/agents.md, a project-scoped file that persists across sessions and loads at the start of each, beside the team's own AGENTS.md. What holds it at Partial: no lifetime, expiry or purge path is published.

Sessions can be resumed by picker or ID, but no memory layer the agent writes and reads across sessions is documented.

Manual mode, the default, asks for confirmation before every mutating action: edits, file operations, terminal commands and MCP tool calls. Plan mode is read-only until the user chooses how the plan is carried out, auto mode is an opt-in, and every turn is a git commit that can be reverted.

The CLI asks for approval before any tool call no allow rule covers, always requires review when switching from Plan to Build, and lets deny rules override allow.

Deployment posture is strong and first-party, and the customer base is highly regulated. What is missing is the other half: no trust center, certification page, attestation, SSO, RBAC or audit control appears anywhere in the site navigation, which for a vendor selling to HSBC, BAE Systems and Lloyds is more likely a disclosure gap than an absence. cosine.sh/air-gapped and cosine.sh/legal are the pages to read.

Strong sandbox, credential and permission controls; the government page states an unnamed authority to operate and IL5 deployability, and no SOC 2, ISO 27001 or FedRAMP attestation has been found.

Checkpointing writes every agent turn into the customer's own git history, which looks like a strong audit trail, but the customer's own systems supply that record, so it does not count for Cosine. What Cosine supplies itself is live visibility, a todo list and a reviewable diff. No vendor-side execution trace, audit log or retained run history is documented.

Every agent action is recorded as a searchable, exportable trajectory of tool calls, file edits, reasoning steps and decisions, propagated to SIEM, with Console dashboards and metrics.

The grade does not depend on Lumen Sovereign, which Cosine labels Coming soon. Air-gapped is a shipped deployment tier with its own solutions page, independent of which model runs inside it.

Full model weights delivered into the customer's own bare metal, VPC, Kubernetes or air-gapped hardware, so no inference path leaves the boundary, with IL5 deployability for government; models are also available managed through Amazon Bedrock.

Swarm mode's specialized subagents are chosen by the orchestrator rather than selected by the customer, so they are Cosine's own machinery, which is Partial. No catalog of prebuilt agents or templates a customer adopts is documented.

Skills, subagents and reusable code-first pipelines are documented, but no browsable library of prebuilt agents or templates.

A published menu of twenty-one models across eight providers, with per-model credit multipliers, is about as explicit as this axis gets.

Per-agent model choice across Poolside's own Malibu, Point and Laguna models and third-party models such as Claude and GPT, with the customer holding full model weights.

No API, SDK, headless mode or MCP server lets an outside caller drive Cosine; the docs name three surfaces, the CLI, Cloud and Desktop. MCP connections let Cosine reach the customer's tools, which is the other direction and counts toward integrations.

External clients drive Poolside agents through the ACP interface (pool acp), alongside first-class MCP hosting, reusable model provider connections, public APIs, availability through Amazon Bedrock and Helm-based deployment; no named public SDK package has been found.

Cosine publishes three internal benchmarks, Niche-Bench, Slop-Bench and Vibe-Bench, with comparative scores against GPT-5.5, Gemini 3.1 Pro and Kimi K2.6. Those measure Cosine's own models rather than giving the customer something to test with, so they do not count here. The grade rests on the agent verifying its own work and on revertible commits.

Agents run the customer's tests and review pull requests, but no harness to evaluate or regression test agent behavior is documented.

Terminal execution, file editing and MCP tool access are all programmatic interfaces, which this axis excludes. The CLI product page, the model menu and the integrations section show no browser, screenshot or GUI automation capability.

Agents work through sandboxes, terminal commands, MCP tool calls, source control APIs and editor extensions; no browser control or screen-based automation is documented.