What AI-Native Companies Actually Look For in Engineers

Main technological themes

AI-native / agentic coding expectations in hiring

• Companies hiring for AI-native development look for engineers who can describe and demonstrate an agentic coding workflow, not just use a single assistant.
• A key red flag: experienced candidates (10+ years) who only use Copilot/one model and can’t explain their workflows beyond that.
• Tools are described as changing quickly (“every 3 months”), so candidates should show continuous hands-on experimentation.

Experiment broadly with “harnesses” (AI coding IDE/agents)

• The speaker lists multiple tools and implies candidates should know them at least somewhat:
  • Cloud Code, Codeex, Cursor, (PI / Pi), Muse, Grok, Build, Open Code
• Red flag: claiming you only ever use Copilot.
• Interview readiness tip: if you claim you use a tool (e.g., Claude Code), you should be able to discuss it in detail.

What to be able to explain in interviews (product/feature analysis)

Models within harnesses (“frontier models”)

• Candidates should know which model they use for which task, with example guidance such as:
  • Fable for hard/complex tasks lasting longer than a few hours and needing the model to keep track without much hand-holding.
  • Opus / Sonnet for orchestration vs. sub-agent execution to balance quality and cost.
• The speaker references token-efficiency tradeoffs:
  • Using Sonnet for implement agents may reduce quality slightly (cited as ~3–4%) but saves tokens (referencing a blog post).

Effort level / inference “difficulty” settings

• Effort level controls speed/cost vs. intelligence/accuracy.
  • Low effort: faster/cheaper, but higher hallucination risk.
  • Higher effort: better accuracy for correctness-critical or complex work.
• Guidance for when to raise effort:
  • Complex/high-risk changes (e.g., refactors affecting 50–100 files) → extra high / max-like effort as needed.

Plan mode vs auto mode (workflow control and safety)

• Plan mode (e.g., in Claude Code)

  • Harness bakes in system prompts to help the agent explore the codebase and requirements before implementing.
  • Recommended for:
    • Clarifying requirements for new features
    • Larger, more complex features
    • High-risk tasks
  • You can also rely on custom planning artifacts (e.g., plan MD files) or use built-in plan mode.

• Auto mode (safety gating / tool execution classifier)

  • An internal classifier sits between the agent and its tool calls.
  • Helps prevent dangerous actions (e.g., attempts like rm -rf should be classified as dangerous and blocked).

• Alternative workflow

  • Sometimes they skip plan mode entirely and use a single prompt with chosen models (e.g., Opus 5.5 or Fable), emphasizing the ability to prompt precisely.

Prompting best practices

• Demonstrates structured prompting:
  • Specify where/what/how (e.g., “left sessions panel… cap at two columns”).
• Mentions productivity through voice input (“Super Whisper”) to speed up iteration—used as an interview advantage.

---

Reusable automation: Skills, agents MD, and “skill hygiene”

Skills (reusable workflows)

• Create a skill when a workflow is repeatable and should be shared across the team.
• Examples mentioned:
  • PR description skill to standardize pull request formatting
  • PR review checklist workflow (risk/security checks)
  • Release skill (tagging/release steps, GitHub actions, etc.)
  • “Deep implementation” skills for multi-step review behavior (to avoid triggering heavy steps every time)

Agents MD / “claude MD”

• Purpose: injected into sessions to enforce consistent agent behavior.
• Contents: bullet-point instructions and “definition of done” items (e.g., “verify tests,” possibly including adversarial reviews).
• Caution:
  • Putting too much into this file can:
    • Increase token usage
    • Slow runs
    • Apply heavy workflows to every prompt (hurting efficiency)
• Recommendation:
  • Move expensive checks into skills used only when needed.

Skill hygiene / MCP hygiene

• Regularly prune unused skills and MCP servers to avoid wasting context window tokens.
• Why it matters:
  • Smaller models with smaller context windows can degrade or get confused by accumulated “trash” context.
  • Even with larger models, unnecessary token consumption is inefficient.

---

Integrations: MCP servers

What MCP servers are

• Described as interfaces exposed by third-party services (e.g., Sentry, Stripe) so a coding agent can connect to external systems.

Example workflows

• Use an Sentry MCP to:
  • Fetch recent error logs
  • Triage issues
  • Potentially create GitHub issues
• Use MCP Stripe to help automate Stripe setup (e.g., create projects/products/descriptions).

Interview expectation

• Candidates should know harness tooling can integrate with third-party services via MCP, rather than manually clicking UIs.

---

Context window and system/tooling understanding

Candidates should understand:

• Context window size and token allocation
• How system prompts/tools/custom agents work
• Memory files and compaction
• How to inspect context usage (e.g., via a “/context” command)
• Identifying MCPs that consume significant token budgets (the speaker mentions “Gmail MCPs” and “fleet”)

---

Agentic depth: parallelization and sandbox/worktree strategies

The speaker emphasizes going beyond “prompt → wait.”

Example approaches:

• For multiple GitHub issues: run implementations in parallel using
  • cloud runners / sandboxes (each issue handled separately)
• Or use local worktrees to work on multiple issues concurrently.

Goal: demonstrate concrete workflow mechanics that improve throughput and efficiency.

---

Interview stance and red-flag guidance

Red flags

If a candidate can’t fluently describe:

• agentic coding workflow
• frontier models
• effort levels
• planning/safety modes
• skills/agents MD
• MCP usage
• context/token hygiene

→ the speaker frames this as a red flag.

Additional stance

• The speaker suggests not interviewing with companies doing “leak code questions,” claiming they may be behind the current industry shift toward agentic tooling.

---

Main speakers / sources

Main speaker

• Unnamed interviewer/coach (speaking throughout the transcript as an interviewer and hiring decision-maker).

Referenced external sources (not direct speakers)

• “Cloud Code” blog post / article about token-vs-quality tradeoffs for model choices.
• Mentions of “Matt PCO’s grill me skill” (referenced as an example skill).