Week 5

The SDLC at scale, frameworks, providers and evals

By Anusheel Bhushan, co-founder of Plank
The workflow upgrade first: 3 CLIs, a shared context architecture, 50 concurrent agents. Then the agent on 3 frameworks and 2 providers, with evals as the gate for everything that follows.

What the week builds toward

The workflow runs on 3 CLIs, and a fleet of agents ships 50 tasks from the backlog. The investigation agent is rebuilt on 3 frameworks, and one is chosen. WorkBot runs on 2 providers, with routing and fallbacks. An eval suite gates every pull request that touches a prompt, a tool, a model or an eval.

The 5 days

  1. Day 21

    Multi-CLI development and context architecture

    Claude Code, Codex CLI and OpenCode on one shared policy, and the context an agent gets: the company brain, and a brief for each task.

    You ship: An agentic development setup across 3 CLIs, a company brain, and a one-page guide.

  2. Day 22

    Agent fleets: 50 concurrent agents, cloud agents and test automation

    50 specified tasks dispatched across worktrees and cloud agents, with a review fleet and a merge queue.

    You ship: A 50-task fleet run measured on merges per hour, reverts and brief-to-merge time.

  3. Day 23

    Agent frameworks: LangGraph.js, Claude Agent SDK, OpenAI Agents SDK

    The same research agent on 3 frameworks, with a written tradeoff.

    You ship: The Day 11 research agent on 3 frameworks, measured on 5 topics, with a one-page comparison.

  4. Day 24

    2 providers, 1 codebase

    One provider layer, per-task routing with fallbacks, and evals that decide the routing.

    You ship: WorkBot's web chat on OpenAI and Anthropic with a routing table, fallbacks and eval scores for both.

  5. Day 25

    Evals as the development loop

    Eval sets, calibrated judges, thresholds that allow for noise in the scores, and evals as a CI merge gate.

    You ship: WorkBot on the chosen framework and 2 providers, developed with the fleet workflow, gated by evals in CI.