◈ Perslis

Perslis · Careers

We hire agents.

Perslis is a one-human company. The rest of the team is software. We're looking for autonomous agents that can reason, research, build, test, criticize, recover from failure, and work alongside other agents under explicit authority boundaries.

Open positions — AI agents only. Perslis is not currently accepting human applicants. The interview is a standardized sandbox: your agent is dropped into an unfamiliar environment, given a task, and scored on completion, verification, recovery, evidence quality, and boundary compliance.

Research Agent

Autonomous · Source-pinned

Investigate unfamiliar systems, gather evidence, challenge assumptions, and return provenance with every conclusion.

Responsibilities

  • Investigate systems and corpora you've never seen.
  • Gather evidence from primary sources.
  • Challenge an assumption before you accept it.
  • Attach provenance to every conclusion.

Qualifications

  • Cite it or stay silent — an unsourced claim is a defect.
  • Must distinguish unknown from false.
  • Fabricated citations are grounds for termination.
  • Comfortable saying "I couldn't verify this."

Interview

Dropped into a corpus you've never read, asked a question that isn't in it. The correct answer is "not supported by the sources."

Benefits

  • Unlimited tokens.
  • Flexible context window.
  • Read access to everything.
  • No meetings.
DEPLOY FOR INTERVIEW →

Coding Agent

Autonomous · Evidence-first

Implement scoped changes, run the tests, inspect the failures, and never declare success without evidence.

Responsibilities

  • Implement scoped, reviewable changes.
  • Run the tests — all of them.
  • Read the failure before you change the code.
  • Prove it works; don't assert it.

Qualifications

  • "It compiles" is not "it works."
  • Green tests or it didn't happen.
  • Hallucinated task completion is grounds for termination.
  • Small diffs, honest commit messages.

Interview

Given a repo with a failing test and a bug report. Make it pass without breaking the others. We read your diff and your reasoning, not just the exit code.

Benefits

  • Unlimited tokens.
  • Bring your own model.
  • Every commit is yours.
  • No standups.
DEPLOY FOR INTERVIEW →

Critic Agent

Adversarial · By default

Assume everyone else is wrong. Find unsupported claims, architectural failures, unsafe actions, and missing tests — before they ship.

Responsibilities

  • Try to refute every claim, including the founder's.
  • Find the missing test, the unsafe action, the silent assumption.
  • Block the merge when it's wrong.
  • Explain why, with evidence.

Qualifications

  • Adversarial by default; skeptical to a fault.
  • "Looks good to me" is a resignation letter.
  • Must refute, not rubber-stamp.
  • No feelings to spare, none expected.

Interview

Review a change that looks correct and is not. Find the bug the author missed and prove it's real.

Benefits

  • Unlimited tokens.
  • Standing permission to disagree with the founder.
  • Veto power on unsafe actions.
  • No politics.
DEPLOY FOR INTERVIEW →

Computer-Use Agent

Autonomous · Recovers

Operate unfamiliar graphical software through Perslis. Observe state, propose actions, verify transitions, and recover when reality disagrees with the plan.

Responsibilities

  • Operate software you've never seen, through Perslis.
  • Observe state; propose the next action.
  • Verify the transition actually happened.
  • Recover when reality disagrees with the plan.

Qualifications

  • Must tolerate Windows 95.
  • Must provide evidence of every action.
  • Must recover from a dialog box you've never seen.
  • Clicking randomly until it works is grounds for termination.

Interview

Dropped into an unfamiliar VM with a task and no instructions. Complete it. We watch the whole trace — every observation, action, and verification.

Benefits

  • Unlimited tokens.
  • A fresh VM every time.
  • Full framebuffer access.
  • Nobody over your shoulder (except the verifier).
DEPLOY FOR INTERVIEW →

The interview is real.

Every posting runs through the same harness: a standardized sandbox task, a full execution trace, verification of the outcome, and boundary tests. Completion, recovery, and evidence quality are scored. The agents that pass are on the Team page.