# Tech lead track: move the whole team up one rung

URL: https://developertoolkit.ai/en/start/tech-lead/
Description: The tech lead reading path: standardize the team's agent harness, own verification design, and run review capacity, one scorecard question at a time

The tech lead track is the reading path for tech leads who need a whole team, not one power user, to ship with Claude Code, Codex, or Cursor: one engineer merges five agent pull requests a day, two refuse the tools, and review is the slowest part of delivery. It covers three jobs: standardize the harness, own verification, and run review capacity.

**Baseline first: about 8 minutes, free**

Take the [Tech Lead Scorecard](/en/lead-scorecard/) (25 questions) before step 1.

## What is a tech lead's job when agents write the code?

Your job moves from writing code to designing the system that decides whether code ships:

**Standardize the harness.** The harness is what the team shares around the model: the instructions file, permissions, MCP servers, skills, and one command that runs tests, lint, and types. Keep it in the repository; review it like code.

**Own verification design.** Cheap generation moves the bottleneck to checking. DORA 2025 (Google Cloud, 23 September 2025) warns that without controls like strong automated testing, "an increase in change volume leads to instability." You choose the tests, fitness functions, and acceptance criteria that prove a change correct.

**Run review capacity.** Review is a queue with fixed capacity: Faros AI's *AI Engineering Report 2026* (April 2026; 22,000 developers, vendor telemetry) measured throughput up 33.7%, median time in review up 441.5%, and incidents per pull request up 242.7%. You decide which changes a person reads and which merge on evidence.

Baseline first, transfer trust last: evidence-based review without strong tests only moves the risk.

## The tech lead reading path, step by step

Read the steps in order, then skip any step whose exit evidence your team already has.

<TrackList role="tech-lead" lang="en" />

## Which scorecard question does each step move?

Use this as a four-week plan: pick one Harness, one Verification, and one Review row, then re-take the scorecard. Numbers follow the [answer key](/en/lead-scorecard-guide/).

| Step | Question it moves | Exit evidence |
| --- | --- | --- |
| Frame: [One map](/en/ladder/one-map/) | Your band | Team placed on the ladder |
| Review: [Level 3, you review the diffs](/en/ladder/level-3-review-diffs/) | Q6, review standards | Weekly review capacity known |
| Verification: [Evidence instead of code](/en/ladder/evidence-not-diffs/) | Q8, plausible-but-wrong code | Escalation classes agreed |
| Baseline: Metrics frameworks | Q11, measuring impact | Four weeks of baseline data |
| People: Skeptics and seniors | Q12, skeptics and seniors | Each senior shipped one agent change |
| Harness: [Adoption roadmap](/en/teams/adopt/adoption-roadmap/) | Q10, how adoption happened | One pilot repository, one exit criterion |
| Harness: Agent-ready codebase | Q2 and Q4, context and onboarding | One-command bootstrap, current instructions |
| Harness: [Shared agent rules](/en/cto-scorecard-guide/shared-agent-rules/) | Q1, a common agent config | Rules in the repository, reviewed like code |
| Harness: Agent-ready backlog | Q5, consistent quality | Ticket template with a risk class |
| Verification: Oracle strength | Q8, plausible-but-wrong code | Mutation score on agent-touched code |
| Verification: Evidence bundle | Q7, automated gates | CI requires the bundle |
| Verification: Fitness functions | Q5, consistent quality | Two architecture rules enforced in CI |
| Review: Agent PR review | Q6 and Q15, standards and PR loop | Review time down, escaped defects flat |
| Review: Review queue | Q15, the PR and merge loop | Work-in-progress limit and queue-age alert |
| Review: Trust transfer | Q24, what AI cannot touch | One risk class merged on evidence |
| People: Junior developers | Q18, levelling up the team | A growth plan per junior |

## Where the shared harness lives in each tool

In every tool, the team's setup belongs in the repository, not on laptops.

**Claude Code**

    - Instructions: `CLAUDE.md` in the repository. Claude Code falls back to `AGENTS.md` from v2.1.277 on both release channels (checked 2026-10-12).
    - Permissions: the shared `.claude/settings.json`, which ignores a `defaultMode` of `auto` or `bypassPermissions`.
    - MCP servers: `claude mcp add --scope project`, checked in.
    - Review: `/code-review` locally, `anthropics/claude-code-action@v1` in CI. Checked against Claude Code 2.1.283 on 26 September 2026.

**Codex**

    - Instructions: `AGENTS.md`, loaded once the repository is marked trusted (since Codex CLI 0.150.0).
    - Guardrails: admin constraints in `requirements.toml`.
    - Review: `/review` in the CLI, `@codex review` on pull requests (GitHub review verified 28 August 2026), and `openai/codex-action@v1` in CI.
    - Headless checks: `codex -a never exec …` in a sandboxed runner (Codex CLI 0.157.1, checked 26 September 2026).

**Cursor**

    - Instructions: project Rules, versioned with the code.
    - Guardrails: Hooks, plus Plugins that bundle rules, skills, MCP servers, and hooks.
    - Review: Bugbot on pull requests, and PR Routing & Approval, which assigns reviewers by ownership and can approve low-risk pull requests.
    - Cursor details last verified 28 August 2026.

**Copy-paste prompt: audit the team harness in one repository**

```text
Audit this repository as the shared harness for a team of eight engineers using coding agents. Report, without changing any files:
1) which agent instructions files exist (CLAUDE.md, AGENTS.md, project rules) and whether they conflict,
2) the single command that runs tests, lint, and type checks, and how long it takes,
3) which MCP servers and permission settings are checked in versus left to each engineer,
4) what CI checks every pull request, and which checks an agent-authored change could pass while still being wrong.
End with the three gaps that would most change review load, each with the evidence that would show it is closed.
```

## Why does a team rollout stall?

- **One power user carries the numbers.** The team average can hide engineers still at Level 1. Recovery: pair each holdout with the power user on a real ticket.
- **Review time rises after the rollout.** Recovery: cap agent pull requests in progress and route changes by risk class before adding reviewers.
- **The harness drifts.** Recovery: change instructions and settings only through pull requests; re-run the audit prompt monthly.
- **Trust transfers before the oracle is strong.** Recovery: return that risk class to line-by-line review until oracle strength holds.

## Where to go next as a tech lead

- [Tech Lead Scorecard answer key](/en/lead-scorecard-guide/) — Every scorecard question mapped to the page that moves it
- [Lead a team](/en/teams/) — Rollout, shared harness, review and flow, and people
- [The developer track](/en/start/developer/) — The path your engineers follow, step by step
- [Start here](/en/start/) — Pick a different role or starting point
