How does coaching an AI agent work?
Coaching an AI agent works like coaching a new employee — except the agent never forgets. You write a plain-English instruction, it becomes a memory block, and it shapes every relevant future cycle automatically.
Coaching an AI agent means writing plain-English instructions that become permanent memory. Unlike prompting a chatbot — where your instruction is only active for that session — coaching an agent creates a persistent memory block that reloads on every future heartbeat cycle. You write it once. The agent applies it indefinitely.
What does a coaching note look like?
A coaching note is a plain-English instruction written the same way you’d brief a new team member:
“Lead with the compliance angle on healthcare CTOs, not ROI.”
“Skip companies with fewer than 10 employees.”
“Subject lines under 6 words get better open rates — keep them short.”
No special syntax. No configuration panel. Just the instruction in natural language.
What happens to the coaching note?
The coaching note is stored as a typed memory block on the agent’s Personal Brain. It’s categorised as a “coaching” type block, assigned a high importance score, and linked to the agent it was written for.
On the next heartbeat, the agent runs Brain Assembly — its context window gets built from top-ranked memory blocks. High-importance coaching notes score near the top and get included. The agent reads the note before deciding what to do. Not as a lookup. As part of how it thinks.
How long does the coaching last?
Permanently, unless you remove or replace it. The memory block doesn’t degrade. It stays in the Personal Brain and keeps surfacing on relevant cycles until someone explicitly changes it.
A single coaching note written today will surface on cycles 30, 60, and 90 days from now — still shaping the agent’s output without anyone repeating it.
Who can write coaching notes?
Two sources:
You: at any time, from the coaching interface. Write the instruction, and it becomes a memory block immediately.
Supervisor agents: In a multi-agent system, director agents (Sales Director, Marketing Director) review worker outputs and write coaching notes automatically. The Sales Director might observe that the SDR’s outreach is too long and write: “Keep first emails under 4 sentences.” That note gets stored and applied on the SDR’s next cycle — no human in the loop.
Can coaching be promoted to the whole team?
Yes. If a coaching insight applies to the whole team — not just one agent — it can be promoted to the Team Brain. Every agent on the team sees it on their next cycle. A pattern observed in the SDR’s outbound (“healthcare buyers respond to compliance framing”) can be promoted to inform the Marketing Manager’s content and the Ad Manager’s targeting.
One observation. No meeting. Everyone updated.