AI in 2027

Your next job is managing AI employees

You may still write code. More of the work can become deciding what to ask for and checking what comes back.

Start with one useful task. You do not need an autonomous team to begin.

Chuck Blake // Developer workflows

Why prepare now?

The tools keep changing.

OpenClaw rebrand + release [2]

Codex desktop app [3]

Gemini 3.1 Pro preview [7]

Claude Design preview [10]

Grok Bot beta [15]

Agent Skills → AAIF [19]

20 dated announcements from 2026, with primary sources. This selection shows changes in what you can try. It does not measure acceleration, reliability or productivity.

How to get started

Ask for a second pair of eyes.

  • Ask AI to review a draft, a plan or a piece of code. You decide which suggestions to use.
  • Talk through an idea. Ask it to challenge assumptions and suggest questions you missed.
  • Use a small example you can share. Keep confidential material out of unapproved tools.

Move at a pace you can handle.

Explore without changing anything

Let AI help you understand the work.

Unfamiliar code

"Explain how this request reaches the database. Point to the files. Tell me what you could not trace."

An architecture decision

"Sketch two ways to add this feature. Compare the trade-offs. Ask me about constraints before recommending one."

Request read-only investigation. Check the explanation against the code and documentation.

A first small change

Choose something you can undo.

  • Refactor one bounded area while keeping its behavior the same.
  • Generate tests for a small function, then check that a deliberately broken version fails.
  • Draft documentation or a worked example, and try the instructions yourself.

Use test data and a review branch. Ask before expanding scope. You can stay with review and explanation as long as they are useful.

Artifacts

Ask for something you can inspect.

The output

A document, diagram, prototype, diff or sample file that stands on its own after the chat ends.

The evidence

References and changed files, checks that actually ran, decisions and trade-offs, and anything still uncertain.

Ask AI to explain what it did and leave work you can review.

A reviewable handoff

Make the explanation part of the task.

"Leave a short walkthrough: what changed, where the files are, why you chose this approach, what you checked, and what I should review next. Flag assumptions and anything you could not verify."

For a CSV export: include a sample file, the test results and a note about empty results or unsupported inputs.

One example of an artifact workflow

Build, inspect, ask for a change.

  • Ask for a small diagram, document or prototype.
  • Open the artifact beside the conversation and inspect it.
  • Point to the part that needs work. Ask for a revision, then review or export the result.

Claude supports this workflow in its Artifacts interface. [27] The habit also works with files and previews in other tools.

Two ways to work

A bionic arm still needs your attention.

Work alongside AI

You ask, inspect the answer and steer the next step. AI extends what you can do while you are there.

Delegate a task

You assign an outcome. An agent uses tools to work through steps, then returns evidence or asks for a decision.

Both are useful. More independence needs clearer boundaries and a way to stop.

What the developer owns

Delegation gives you a management job.

Choose

Pick work worth handing off.

Brief

Describe the outcome and context.

Bound

Limit access, scope and cost.

Evaluate

Check the result.

Accept

Decide what ships.

Follow up

Keep an owner for the outcome.

Agents have no human accountability. You and your organization still do.

Choose

Start where you can check the answer.

  • Pick a clear finish and a result you know how to evaluate.
  • Prefer a small, reversible change over an open-ended rewrite.
  • Count review and corrections as part of the effort. More generated code is not the goal.

Example for the rest of the talk: add a CSV export to an invoice list, using test data and a review branch.

Brief

Give the agent a brief it can work from.

Outcome and context

Let an authorized user export the invoices matching the current filters. Follow the existing export pattern and permission checks.

Definition of done

Return a reviewable diff, tests and a sample CSV from fixtures. Explain decisions, unsupported cases and anything you could not verify.

Link relevant files and a known-good example. Ask the agent to surface missing decisions before implementation.

Boundaries and human control

Give it only the access this task needs.

  • Use a review branch and test data. Keep production credentials out.
  • Set a time or spend limit. Stop at the limit or when the task needs a new decision.
  • Require approval before publishing, spending money, deleting data or changing permissions.
  • Keep a practical way to pause the work and revoke access.

A prompt describes the boundary. Tool permissions enforce it.

Evaluate

Write the checks before you see the answer.

  • Does the sample contain the right filtered rows, including an empty result?
  • Can a user access another customer’s invoices? Check the denial path.
  • Do quoting, line breaks and spreadsheet-formula inputs behave safely?
  • Try the download flow yourself. Read the artifact and the changes.

A second agent can help review. Agreement between agents is not proof.

Accept

"Done" is a claim you have to check.

What AI returns

The output, references, actual check results and a short account of risks or unverified behavior.

What you decide

Inspect the change. Run your checks. Accept it, ask for a specific correction or stop the work.

You do not need to type every line to be responsible for what ships.

Practices worth borrowing

Learn one practice before adopting a whole system.

Matt Pocock / skills

Try a grilling session to clarify a plan, or a test-driven loop for one small change. [23]

Poteto / pstack

Try a read-only investigation, a small prototype or a bug fix with runtime evidence. [25]

These are examples you can read and adapt. You do not need to install either collection to begin.

What employers need to provide

Make room for practice and review.

  • Provide approved tools, clear data rules and a safe place to experiment.
  • Pair beginners with someone who can help check the work. Share failures as well as useful prompts.
  • Allow time to learn. Review and corrections are part of the workload.
  • Judge useful outcomes, defects and total effort, rather than lines generated or agents launched.
A practice plan you can slow down

Try one useful workflow, then decide.

First

Ask for a review or explanation. Save what helped and what was wrong.

Next

Request a small artifact. Revise it together and check its references.

When ready

Delegate one reversible change with a clear boundary and a review checkpoint.

Then

Repeat, adjust or stop based on quality and your total effort.

Stay at any step as long as you need. You can get useful work done without running AI unattended.

Your next experiment

Which task would you try tomorrow?

Pick something small enough to check. Ask for an artifact that explains its work.

Resources, starter prompts and the sourced 2026 timeline →

Audience view hides speaker notes. Use the Notes button or open the speaker view. Notes are public, not private.

Chuck Blake // chuckblake.com