~/talk

Context Engineering for MCP: 150K-22K Tokens

links:none on file
name:Sonam Raj
links:none on file
slides:none on file

Measured context goes from 1.17M tokens for 2,594 tools down to a single-code approach that often needs ~1k tokens; a production agent moves from 150K to 22K prompt tokens and tool selection jumps from 42% to 80%. Models write JavaScript in a secure sandbox to call MCP tools, returning trimmed fields instead of full documents; prototypes (Pi demo) work but face stale definitions, nonexistent method guesses, excessive script output, and harder debugging.

AI-written summary · about

  1. 0:00The 1.17 Million Token Problem
  2. 1:01Meet the Speakers and Takeaways
  3. 2:30Where Did the Context Go
  4. 3:44Copying Data Through the Model
  5. 4:31Tool Selection Accuracy at 42%
  6. 4:56Tool Selection Accuracy at 42%
  7. 5:40How Code Mode Works
  8. 8:25Progressive Disclosure of Tools
  9. 9:05Production Results: 150K to 22K
  10. 10:53Pi Demo: Direct Tool Calling
  11. 12:57Pi Demo: Code Mode Run
  12. 16:38Challenges and When Not to Use It
  13. 19:01Try It Yourself and Wrap-Up
  14. 19:52Try It Yourself
~/search
type to search 6,518 conferences, 143,918 speakers and 192,133 talks
~/help
cmdctrl k
search every conference, speaker and series
/
focus the filter search
h
home
c
conferences
s
speakers
e
series
t
light / dark — the status bar button does the same
← →
on a talk: 5 seconds back / forward
0–9
on a talk: jump to 0%–90% of the recording
?
this help
esc
close this, then clear every filter — or just the search, while you are typing in it

Ignored while you are typing. Pages without a filter rail have no search to focus.