We need to build agents that get better every run. We need the same harness rules for every agent, whether it's reshaping a fellow's portal or running outreach.
## 1. Real self-recursive loops
Every agent runs: ==observe → decide → act → verify → remember → (maybe) retry==.
- Cap the loop (6–12 steps). Infinite "keep thinking" is a bug.
- Retry only when there's a verifier (send confirmed, CRM write landed, schema passed, citation found). Reflection without a check is theatre.
- Persist the plan before spawning sub-agents - a crash shouldn't lose the thread.
- Always allow a quiet path: nothing useful to do → do nothing.
## 2. Auto-research, grounded
- Cheap checks / change-detection first; spend on deep research only when something new shows up.
- Cite before you show. No claim or "comparable company" without a source. Can't cite it? Send to studio review, don't show the fellow.
- Keep research isolated — it runs in its own context and returns a short, verified brief. Don't let a research dump pollute the message you send.
## 3. Memory in four layers (not "the chat")
- **Working** — this run only
- **Episodic** — what we did + what happened (source of truth)
- **Semantic** — durable facts (ICP, stage, tone, prefs)
- **Procedural** — playbooks/skills (human-owned)
Garbled summary? Rebuild from the activity log + raw facts. Agents can update notes on a fellow; they must not silently rewrite shared playbooks.
## 4. Skills/helpers as inventory
Every helper ships with a named owner, a playbook (skill), tests, and a trust level (supervised → semi-trusted → independent). One studio registry for all of them — no per-product inventories.
## 5. Human-in-the-loop by default for anything touching people or money
- Small, reversible tweaks can be automatic (e.g. reorder a layout module).
- Bigger changes, any message to a fellow, pricing/legal, and outbound sends → studio approval first.
- Approvals must be resumable — pause the workflow, don't lose state if a tab closes.
## 6. Cost + blast-radius discipline
- Cheap models/rules for routine work; strong models only for hard judgment and fellow-facing copy.
- Log every billable step.
- Prefer flipping flags/settings over auto-writing prod code. Specs and PRs are fine; silent codegen into prod is not — not for pilot.
## 7. Compounding is the point
Production traces → eval cases → improve the skill → next fellow benefits.