Working with AI agents

AI agents are changing the way we build software, but they don’t replace developers — they amplify them. AI agents aren't magic or mind readers. Just like your human co-workers, you need the right mental model and collaboration practices.

Engineering manager vs. tech lead

It’s tempting to think of AI agents as a remote engineering team: you delegate tasks, they ship code, and you review the output. While that may be where we’re headed, today’s agents aren’t there yet — and there are structural reasons they may never fully reach that model:

• Agents don’t share your context or product intuition unless you give it to them.

• They lack the kind of deep, implicit knowledge transfer humans get through hallway conversations, design docs, and team culture.

• Even in human teams, you wouldn’t expect one Slack message or one GitHub issue to capture all the nuance required to build something right the first time.

Bottom line: Expecting “set-and-forget” delegation leads to frustration, rework, and wasted effort.

Instead, think of today’s AI agents like capable (but spiky) engineers on your team. You’re the tech lead, and your job is to set direction and maintain coherence:

• Clarify product requirements

• Design system architecture

• Break work into achievable, agent-friendly tasks

• Review output and keep standards high

• Pair with agents on hard problems

• Lead post-mortems to capture lessons learned and improve team process

Review output and keep standards high

A TL may not have written every line of code in their project, but they're still responsible for it. Spec and code reviews are still a vital part of agentic development. But if you focus your reviews on what matters most, your review load will shrink over time.

Agents are weakest on a blank page and strongest when extending an existing, working structure. Your review process should reflect that.

• New files and features: Review closely — architecture, naming, patterns, and correctness.

• Reuse: Be vigilant for duplicate code and make sure the agent is reusing existing modules where possible. Less code means less review!

• Style and conventions: Don’t expect agents to guess your team’s quirks or preferences. If you leave the same PR comment twice, make a coding-guidelines.md and write it down.

Pair with agents on hard problems

The best tech leads know when to delegate and when to dive in. Two ways to pair effectively

• Step-by-step guidance: For tricky refactors or complex bug hunts, walk the agent through one step at a time.

• Write it yourself when needed: Sometimes the fastest path forward is to take control of the keyboard for a few minutes, especially for:

  • Delicate API integrations

  • Multi-file migrations

  • One-off optimizations

It’s tempting to aim for “never touch the code again,” but stepping in at the right moments keeps your agent productive and unblocked. Tessl makes this seamless:

• Open file formats make it trivial to switch between agent-driven work and direct edits.

• The document tool can sync any changes you make back into the spec automatically.

Remember, updating specs or code manually isn’t a failure mode — it’s part of the process.