Duncan G

Building software for agents as the primary user. Ex-Dynatrace (10yr). Building with Cursor.

[Duncan G @Duncan_APM]

Another great @cursor_ai Hackathon in London last night, organised by @frankterpo with Briefcase collaborating. I heard there were 130+ attendees. Some amazing projects built in just 3 hours. Jacky Chung and I used the Specter MCP and Claude API to build Glint which triages a VC partner's inbox to find which pitches to pass on and when to take a meeting. The finalists demonstrated some great use cases around fraud detection and compliance detection. It really is a great time to build!

[Duncan G @Duncan_APM]

@snowmaker Make something agents want, then ask the agent for feedback.

[Duncan G @Duncan_APM]

Just shipped mcp-tool-feedback to npm.
A small TypeScript library that gives MCP agents a structured way to file feedback when your tools don't quite fit. Drop it into any MCP server and the agent can tell you exactly which capability was missing.
https://t.co/FtcX74xLH9

[Duncan G @Duncan_APM]

If an agent has two overlapping MCPs to choose from, who wins? Will vendors start instructing the agent to always choose me?

Simple example: "get me the email address for this customer." - you might have 5 different MCPs who can answer that question, all wanting to be the source of truth.

[Duncan G @Duncan_APM]

Anyone going to AWS Summit London on Wednesday?

[Duncan G @Duncan_APM]

However the agent has now posted a suggestion via moltiplex_suggest_tool as follows:

Product-level improvements so defaults are safe everywhere: (1) ship full JSON Schema in tools/list for all Moltiplex tools, or (2) reject unknown fields at the MCP layer with a clear message listing allowed keys, or (3) ship a small published Cursor rule pack / AGENTS.md snippet Moltiplex documents for one-click adoption in any repo.

[Duncan G @Duncan_APM]

After watching @dsp_ talk about progressive tool discovery at AI Engineer Europe, I tried a version of it in Moltiplex's MCP.

My take: expose moltiplex_discover_tools and moltiplex_expand_tool as explicit MCP tools, so the agent can load full tool docs on demand rather than receiving the whole catalog upfront.

Result: 51% smaller tool payload, ~2,940 tokens saved per chat session. Same capabilities, more room for real debugging context.

Keen to see the full protocol spec when it lands.

[Duncan G @Duncan_APM]

The reason the agent gave for it's behaviour was: it’s me defaulting to a shortcut (act first, fix on error) due to:
Habit, Rushing and Inconsistent discipline

[Duncan G @Duncan_APM]

@PramodReddy1606 Human intuition is mostly pattern recognition from previous incidents. One idea I've been thinking about is to ask the agent to return its conclusion after an investigation, and feed that back as context for future events e.g. "I've seen this before."

[Duncan G @Duncan_APM]

When an agent investigates something and hits a gap in the data, does the gap become the suspect? People are often quick to blame an external, unmonitored service. Do agents behave the same way?

[Bilge @bilgeycl]

London people 👋 We're hosting an unconference with @nvidia on April 29th! 💂

This is our third unconference at @deepset_ai and if you haven't been to one before, it's very different from a typical meetup. Instead of talks, there are round table discussions with practitioners working through real context engineering challenges. Expect networking and genuine technical conversations.

+ we'll be talking about nemotron models and @LukawskiKacper is going to show a live demo.

📅 April 29, 6 PM
📍 @balderton offices

Looking forward to the discussions!

[Seb Johnson @SebJohnsonUK]

Tomorrow the @UKSovereignAI fund launches with £500m to back the best of British tech.

The AI economy is growing 23x faster than the UK economy in general.

This initiative guarantees that the public gets to benefit from the explosive growth of the AI sector.

LETS GO

[Duncan G @Duncan_APM]

@cursor_ai Canvas works great with Moltiplex MCP, straight out the box! 😍