How AI agents build & ship eBuzz
A fleet of 31 specialised build agents writes, tests, secures and deploys the platform. A small human team supervises and approves every gate that matters: architecture, security, money, legal and production releases. It takes 26 weeks from kickoff to public launch in the wedge vertical.
Timeline (26 weeks)
Phases, agents and milestone gates (◆). Bar colour shows the main model each agent uses.
How agents build & deploy, end to end
Every change moves through the same governed pipeline. Agents do the work, and people approve the steps that carry risk.
- No production data or credentials, ever
- One scoped GitHub token per agent; no direct pushes to main
- Network egress allow-list (package registries, vendor sandboxes)
- Token and time budgets per ticket; the Orchestrator kills runaway loops
- Full transcript and tool-call audit trail in Langfuse
- Architecture & threat models (CTO, Security Lead)
- Protected-path PRs: two-person rule
- Legal documents (outside counsel)
- Every production release touching money or auth
- Go/no-go at week 22 and launch
- Acceptance criteria pass in e2e tests
- No new critical/high security findings
- Accessibility checks pass
- Docs & runbooks updated
- Agent evals ≥ threshold (for AI features)
Detailed plan by agent
Name, purpose, goal, tasks, deliverables, dependencies, human gate, estimated tokens, time and cost. Click an agent to expand it.
Token usage & cost model
Estimates cover all tokens the agents process, including cached context re-reads. Change the assumptions and every number on this page recalculates.
Prices are Anthropic first-party list prices per million tokens as of September 2026. Cache writes are charged at 1.25× input. Check current pricing before budgeting.
Human team (supervising the agents)
About 10.5 FTE over 26 weeks. A traditional team building the same scope would need roughly 35–45 engineers, designers and QA staff.