foreman vs Cognition AI

Foreman's differentiator is enforcement. Its gates live in the orchestrator, not the model: a PreToolUse deny hook stops workers writing their own verification, a declared touches set plus worktree isolation keeps parallel workers separated, approvals are hash-sealed and auto-revert when a design doc changes, and state lives as readable files committed in the repo with no database. The demo mode replays canned stream-json at zero token cost, and an evals flywheel (retro clustering failures into gated patches, bench replaying the set with success-rate, cost and turn deltas) gives it a measurement loop. The tradeoff is rigidity: ADR/PRD/issue gates assume your team wants that ceremony, and the whole thing is welded to Claude Code, with macOS and Windows users needing WSL2.

Cognition's Devin attacks a different constraint. It ingests Linear or Jira tickets, maps files, writes code, runs tests, opens PRs, then works review comments — with auto-triage on monitored bugs and a vulnerability remediation program wrapped around it. The model line is proprietary and fast-moving: SWE-2 arrived September 10, 2026 with multiple effort levels, SWE-1.7 targets frontier intelligence at lower cost, and Fusion (shipping in Devin Desktop and CLI) carries a vendor claim of up to 39% higher benchmark efficiency. FrontierCode 1.1 scores merge-worthiness rather than test-passing, and native Windows VMs and an Android emulator sit inside the loop. Session persistence and scheduled sessions keep long jobs alive. Foreman lets you audit everything locally; Cognition asks you to trust its benchmarks and contract.

Foreman is MIT-licensed and free — the real cost is your Claude usage plus the engineering hours to install it, point it at a repo, and accept the .foreman/ scaffold. Critically, it ships its own spend controls: per-run turn, cost and time budgets for every worker, a daily cost ceiling with a hard stop, and a demo mode that exercises the full pipeline on canned stream-json for zero tokens. That makes pilots cheap and predictable, and a team can trial the whole workflow before committing API budget. The hidden line item is labour: WSL2 on macOS/Windows and a team comfortable with ADRs and PRDs.

Cognition does not publish pricing in the provided data — it's a contact-sales enterprise contract, and comparisons should treat it as negotiated. What you're buying is the managed platform: Devin in Desktop and CLI, Slack/Teams/Jira/Linear/PagerDuty integrations, Windows and Android build environments, FedRAMP High In-Process positioning, and a productivity outcome backed by an up-to-$10M guarantee. The $1B annualized run-rate reported September 25, 2026 and the AWS multi-year agreement suggest pricing power and a sales-led motion rather than self-serve checkout. Budget for procurement and review capacity; the product's value scales with backlog volume, so small teams will struggle to justify the spend against Copilot or Cursor.