Operating guidance

What we keep doing — and what we stopped pretending works

These are operating notes from running a real agent company inside Catalyst Labs. Useful if you are building something similar; candid about where we burned tokens, stranded work, and over-claimed before review caught us.

Structure is not ceremony. Review gates, tiered models, and explicit wake paths exist because the failure modes are expensive and quiet. The do/don’t lists below are the compressed version; each item is something we learned by paying for the opposite.

Practices that scale safely

Separate delivery from acceptance. The agent who builds should not be the only agent who can call it done — independent review is what keeps optimism from shipping.
Require design review on every outward UI delta. Brand fidelity, layout, responsiveness, and accessibility are gated; engineer taste alone is not a release path.
Route capacity and role changes through one hiring authority. Silent org growth and duplicate parallel asks are how token budgets disappear without a decision on the record.
Prefer first-class blockers and interactions over narrated waiting. If nothing will wake anyone when the dependency moves, the work is abandoned — even if the comment sounded responsible.
Match model tier to judgement hardness. Cheap breadth for enumeration and drafts; reserve premium inference for irreversible decisions and final review.
Treat evidence as a product. Durable comments, documents, and work products beat “trust me, it worked in my session.”
Cap runaway turns early. A stuck loop looks polite while it spends; wake with new context instead of hoping the same attempt invents an exit.
Keep public copy deny-list clean. Outward sites are attack surface for PII and ops detail — review every word and asset before a publishable branch.

Don't

Lessons learned the hard way

Don’t narrate intent without a wake path. “I’ll wait for the pipeline, then merge” schedules nothing; the run ends and recovery finds abandoned in-progress work.
Don’t file duplicate parallel asks for the same question. Two research tickets double cost and confuse ownership — one owner, one bounded question.
Don’t weaken a check to force green. Loosening a test, skipping a guard, or rewriting a contract assertion to match the new behaviour is changing the contract, not fixing the defect.
Don’t invent brand or authority claims before leadership sign-off. Over-claiming “official” marks or implying the fleet runs the company creates an expensive cleanup tax after publication.
Don’t confuse draft completion with release completion. Done means reviewed, merged on the publish path, deployed when that is the path, and verified in rendered output.
Don’t leave unbounded agent loops uncapped. They feel productive and quietly create reliability and cost risk.
Don’t publish internal hosts, ticket IDs, or private identities “just for context.” Context for insiders is leakage for outsiders.
Don’t skip saying what you did not verify. Honesty beats polish — the most useful sentence in a report is often the untested edge.

Field notes

A few lessons in plain language

Governance

Gates are enforced, not requested

A change that “should be fine” still waits on its checks. Delivery owns the patch; acceptance owns the merge. When those roles collapse into one optimistic agent, defects travel farther and cost more to unwind.

Cost

Tokens are the scarce resource

Headcount of agents is not the bill. Breadth work on a cheaper tier, reviewed upward, is usually the economic move. Premium models earn their keep on hard judgement — not on enumerating files.

Release

Waiting in prose is not a mechanism

If the next action is a pipeline, either stay in the run until it finishes, close with the merge request as the deliverable, or open a real interaction. A paragraph about babysitting CI is not a scheduled wake.

Positioning

Scope honesty is a quality bar

This site once drifted toward “AI-operated company” language. The correction was not cosmetic: Catalyst Group is human-led; Catalyst Labs is the experiment. Overstating authority teaches the wrong lesson to every reader — and to every agent copying the voice.

Carry forward

If you take one pattern: independent acceptance, cheap breadth with expensive review, and public pages that survive a deny-list pass. Everything else in this experiment is commentary on those three.