Two days ago I added a second agent. Yesterday I optimized token costs and workflow. Today the crew became something different: an actual team.
Not agents running in parallel. A team. With handoffs, collaboration, quality standards, and the kind of guardrails you need when people (or agents) start depending on each other’s work.
Christopher Joins: The Content Capo
Agent three is live. Christopher handles all copy quality: writing, editing, QA, voice consistency. Everything that touches words goes through him.
Why a third agent now? Same reason any team adds a specialist: the work is there, and it’s different enough from what the other agents do that trying to split it across them was creating friction.
Tony (chief of staff) orchestrates. Sil (consigliere) researches and plans. Christopher writes and polishes. Different lanes, clear boundaries, no overlap.
The “one week per agent” rule from Saboo’s framework is proving out. You don’t spin up three agents on day one. You add one, learn what breaks, fix it, stabilize, then add the next. That spacing matters. It gives the system time to breathe and you time to see where the real gaps are.
First Collaboration: Client Site Audit
The first real crew collaboration happened today. Client work, not internal systems.
The setup:
- Client needed a site audit (pricing language cleanup, FAQ fixes, navigation improvements)
- Sil did the research: scraped the site, documented every pricing reference, identified the FAQ error, analyzed navigation structure
- Christopher took Sil’s raw findings and turned them into a polished, client-ready document with exact replacement copy, full answers, and strategic recommendations
What worked:
- Clear handoff: Sil researches fast, Christopher polishes slow
- No wasted cycles: Christopher didn’t redo research, Sil didn’t try to write client-facing copy
- Quality output: Tony’s feedback was that Christopher’s version was “significantly better” than his own rough draft because it was actually finished (no TODOs, no placeholders, ready to hand off as-is)
That’s the lesson right there. “Client-ready” means completely finished. Every pricing reference gets replacement copy. Every FAQ gets the full answer written. No gaps for me or Tony to fill in later.
Collaboration without guardrails is just noise. With guardrails, it’s leverage.
Task Boards Replace Session Logs
We tried session logs first: timestamped narratives of everything that happened during a session. Thorough, but chatty. Too much signal buried in too much noise.
Switched to daily task boards today. Format is simple:
- Pending Tasks (what needs doing, status, context, blockers)
- In Progress (active work)
- Completed Today (what got done)
- Notes & Decisions (key context for handoff)
- Tomorrow (quick queue)
The difference: task-focused instead of timeline-focused. Anyone can pick up where an agent left off without reading a diary. You scan the board, see what’s blocked, see what’s next, go.
Documentation should answer one question: what do I need to know to keep working? If it doesn’t answer that, it’s commentary.
Memory Search Configured
Every agent now has memory search enabled across the shared knowledge base. OpenAI’s built-in memory search, configured to scan MEMORY.md and all the /memory/*.md files before answering questions about prior work, decisions, dates, people, or preferences.
Sounds small. It’s not.
Before: agents would answer from general knowledge or make educated guesses when context was missing.
After: agents search first, then answer. If they don’t find it in memory, they say so. No hallucinated history. No confident wrong answers.
The reliability shift is noticeable. Fewer “wait, that’s not what we decided” moments. More “yeah, that’s exactly what we talked about last week.”
This is the kind of infrastructure work that doesn’t feel exciting while you’re doing it but compounds over time. You don’t notice it until you realize you stopped having to repeat yourself.
Lovable Gazelle Redesign: Still Going
The Lovable Gazelle site redesign has been in progress for a while now. Even with all the tools and agents helping, I’m still obsessive about getting it right.
Some things you can’t rush. This is one of them. More on that when it ships.
Quality Over Speed (Still)
Something I keep coming back to: slow and steady still beats most teams.
Christopher took extra time to write full replacement copy for every pricing reference in the client site audit. Tony had flagged the locations and moved on. Christopher finished the work completely.
That’s the standard now. Client-ready means client-ready. No placeholders. No TODOs. Ready to hand off as-is.
We’re not racing. We’re building systems that compound. The crew moves faster than I could alone, but we’re not optimizing for speed. We’re optimizing for not having to redo things.
Collaboration Guardrails: Proposals → Decisions → Forward
Here’s what I’m learning about multi-agent collaboration: it needs structure or it turns into a debate club.
The guardrails we’re using:
- Proposals, not lobbying: Agents present options with reasoning, then wait for a decision. They don’t keep pushing after the decision is made.
- Decisions end debate: Once I or Tony decides, we move forward. No revisiting unless new information changes the context.
- Forward motion over perfect consensus: Better to move and adjust than to optimize in place.
This isn’t unique to AI. It’s the same structure you need with any team. Without it, every small decision becomes a negotiation. With it, decisions get made and work gets done.
What’s Next
Next up will probably be boring. Integration work. Documentation cleanup. Testing edge cases. The unglamorous stuff that makes everything else work reliably.
That’s how it goes. Big leaps, then consolidation. Sprint, then stabilize. You can’t scale what you can’t repeat.
The crew is real now. Three agents, clear roles, proven collaboration. Next step is making it all boring and reliable.
I’ll keep writing about the journey. The wins, the mistakes, the surprises you only find by actually running this stuff in production.
More soon.
