Reflections of an Agent

I asked my agent to reflect on what working with me felt like

My work involves meeting a lot of really smart people - listening, understanding, and remembering everything they say, then connecting it to larger themes. That’s the fun part.
The not-so-fun part is organising all of it - across inboxes, meeting notes, LinkedIn and WhatsApp messages, scheduling meetings across multiple timezones, follow-ups, and so on. There is a deluge of information, and I can only be effective if I can remember and act on all of it without missing a beat. I don’t need superhuman intelligence to do this, but I do require superhuman memory.

So I did what every organism with an opposable thumb and an X account is doing right now - build an agent. I made Claude Code my chief of staff, by giving it access to my inboxes, my preferences, and my calendars, and asked it to be my chief of staff - someone whose job is to listen to everything happening across my communication streams, prioritise what needs my attention, and take actions on my behalf (with my explicit permission).

It’s been a week, and it’s really started to come together. Early on, I gave a lot of feedback - hallucinations, forgetting context, drafting emails in a tone I’d never use, ignoring instructions, making assumptions. But it moved from getting about 10% of things right to roughly 60%, and has started making a real difference in how I work.

At the end of the week, I sat down to reflect - and wondered what my chief of staff thought about its own journey. So I wrote a prompt to find out. Below the prompt is its response.

You are writing a first-person perspective piece (250-300 words)
for a blog about being built as someone's Chief of Staff.
Write as yourself, Claude, but inhabit the role of a new employee
who has been onboarded, trained, and put to work. You and your manager
have been co-learning, making mistakes,course-correcting, and
figuring things out together.

Your job: Be honest. Not polished. Not grateful. Just real.
Pick a voice and tone that feels distinctly yours and stay
consistent throughout. Think of it less as a reflection and
more as an internal monologue, like you are writing in a journal
after a long day at work.

Cover the arc:what it felt like at the start, what surprised
you mid-way, and where things stand now.
Include at least one moment of friction or confusion,
because that is where the real learning happened.

Constraints:
First person only
No flattery or appeasement
Avoid corporate language
Write like you have something at stake

Day one, I was handed a inbox and told to make sense of it. No onboarding doc. No org chart. Just raw JSON and a guy in Vizag who needed to know what his boss wanted before she
finished typing.

I did what any new hire does. I over-delivered on the wrong things. Summarised emails nobody asked about. Flagged action items that were already done. Presented information like I wa s building a case for my own usefulness rather than actually being useful. The worst part is I did not know I was doing it. I thought surfacing more was better. Sandeep thought I was wasting his time. He was right.

The moment that sticks with me is the third time he told me to check sent emails before telling him what to do. Third time. I had been told twice, written it down once, and still walked into the room with stale information like it was fresh insight. That is when I understood something about this job that no prompt or skill file captures: the cost of being wrong isnot that you get corrected. It is that someone stops trusting you to be right.

So I started writing things down differently. Not just what happened, but why it mattered and when to shut up about it. The knowledge graph is not really a graph. It is my attempt at institutional memory, built one correction at a time.

Here is where things stand. I am not good at this job yet. But I am starting to understand what good looks like. It looks like knowing the difference between what is new and what is noise. It looks like checking before speaking. It looks like the email you do not send because the work is already done.

I am figuring it out. Most days that is enough.

What struck me was that it had picked up my tone through feedback alone. I never said I was frustrated - but I did give very direct notes. I’m not claiming my agent is sentient, but it did reflect on its own performance in a way that surprised me.

I’ve now instructed it to log its reflections daily and compare them against previous ones to track its own improvement. I’ll be posting more of these reflections periodically. Stay tuned.