BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//HasGeek//NONSGML Funnel//EN
DESCRIPTION:The Fifth Elephant - Hyderabad meetup
X-WR-CALDESC:The Fifth Elephant - Hyderabad meetup
NAME:Agent harnesses & small models in production
X-WR-CALNAME:Agent harnesses & small models in production
REFRESH-INTERVAL;VALUE=DURATION:PT12H
SUMMARY:Agent harnesses & small models in production
TIMEZONE-ID:Asia/Kolkata
X-PUBLISHED-TTL:PT12H
X-WR-TIMEZONE:Asia/Kolkata
BEGIN:VEVENT
SUMMARY:Check-in and introductions
DTSTART:20260905T041500Z
DTEND:20260905T044000Z
DTSTAMP:20260911T054458Z
UID:session/LWqWkHkN2u4VyfSgZQLRTs@hasgeek.com
SEQUENCE:3
CREATED:20260901T041920Z
LAST-MODIFIED:20260901T043201Z
LOCATION:Hyderabad
ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com
BEGIN:VALARM
ACTION:display
DESCRIPTION:Check-in and introductions in 5 minutes
TRIGGER:-PT5M
END:VALARM
END:VEVENT
BEGIN:VEVENT
SUMMARY:Introduction to the meet-up & reviewers
DTSTART:20260905T044000Z
DTEND:20260905T045000Z
DTSTAMP:20260911T054458Z
UID:session/VCg8k7DRyTN7NpkhkyhYxF@hasgeek.com
SEQUENCE:3
CREATED:20260901T041939Z
LAST-MODIFIED:20260901T043221Z
LOCATION:Hyderabad
ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com
BEGIN:VALARM
ACTION:display
DESCRIPTION:Introduction to the meet-up & reviewers in 5 minutes
TRIGGER:-PT5M
END:VALARM
END:VEVENT
BEGIN:VEVENT
SUMMARY:Who watches the coding agent? Building harnesses that actually wor
 k.
DTSTART:20260905T045000Z
DTEND:20260905T052500Z
DTSTAMP:20260911T054458Z
UID:session/5QPEUZSt6E1hmrxNvuinKP@hasgeek.com
SEQUENCE:4
CATEGORIES:30 mins talk
CREATED:20260901T042238Z
DESCRIPTION:*Instrumenting your coding agent's harness against silent fail
 ure*\n\nMost teams that build a coding-agent harness (a `CLAUDE.md`\, a fe
 w hooks\, a CI gate) treat it as config they set once. It isn't. The harne
 ss is supposed to guard your code changes. Those same changes erode the ha
 rness itself: the standing-rules file goes stale until it's confidently wr
 ong rather than usefully absent\, rule density crosses an adherence cliff 
 where even frontier models stop following it\, and the agent doing the wor
 k can satisfy a failing gate by deleting the test that enforces it. Nobody
 's watching the watcher\, and a harness that fails silently is worse than 
 no harness at all.\n\nThis talk gives you a failure taxonomy for the harne
 ss itself and the safety net for each failure. The core distinction: **enc
 oded guardrails** (lint\, tests\, static analysis\, anything the agent doi
 ng the work can edit or delete) versus **structural guardrails** (permissi
 on hooks the agent's own session can't switch off\, default-deny network e
 gress\, a CI gate with no admin-bypass path). The first kind only holds if
  the agent cooperates\; the second holds regardless of what the agent trie
 s. "Coverage is theater" generalizes: every encoded guardrail is theater a
 gainst a suggestible actor. \n\nFrom there\, three concrete practices for 
 treating the harness as production infrastructure rather than settings: co
 nfig-drift detection in CI\, canary load-lines that confirm context actual
 ly loaded\, and harness regression evals that gate every change to a promp
 t\, tool description\, or iteration cap before it ships.\n# Takeaways\n1. 
 Your harness has its own failure modes (context rot\, adherence cliffs\, g
 uardrail erosion) and needs sensors pointed at itself\, not just at the co
 de it guards.\n2. One question separates a real guardrail from a decorativ
 e one: can the agent doing the work delete this and still pass? If yes\, y
 ou're relying on an encoded check\, not a structural one. Know which you h
 ave before you bet on it.\n# Audience\nEngineers and teams running coding 
 agents against real production codebases who've already written a `CLAUDE.
 md`/`AGENTS.md` or a hook and want to know whether it's actually holding. 
 Includes people building agent infrastructure\, harness tooling\, and eval
 uation frameworks.\n# Bio\n[Karun Japhet](https://karun.me) is an engineer
  and consultant at [Sahaj Software](https://sahaj.ai)\, working on large-s
 cale systems and distributed architecture. For the last two years his focu
 s has shifted to how AI reshapes the full software development lifecycle\,
  with a particular interest in harness design (the scaffolding and guardra
 ils around coding agents\, not just the prompts fed to them). Karun writes
  and speaks on codebase structuring for multi-agent reasoning\, measuring 
 AI-assisted development impact\, and how much judgment to leave with human
 s as agents take on more of the work. He's spoken previously on distribute
 d systems\, event-driven architecture\, and AI's role across the delivery 
 lifecycle.\n# Others\n**Slides link (PDF\, comments access):** TODO\n\n**2
 -min elevator pitch video:** TODO\n\n**Past recorded talks:** [YouTube Pla
 ylist](https://www.youtube.com/playlist?list=PLY67XcOB0u1SCXn5Z8ZcMnDJHWV4
 DuPvV)\n
LAST-MODIFIED:20260901T043204Z
LOCATION:Hyderabad
ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com
URL:https://hasgeek.com/fifthelephant/agent-harnesses-small-models-hyd-mee
 tup/schedule/a-green-harness-is-not-a-healthy-harness-5QPEUZSt6E1hmrxNvuin
 KP
BEGIN:VALARM
ACTION:display
DESCRIPTION:Who watches the coding agent? Building harnesses that actually
  work. in 5 minutes
TRIGGER:-PT5M
END:VALARM
END:VEVENT
BEGIN:VEVENT
SUMMARY:Benchmarks that lie: when “good enough” models fail the thing 
 that matters
DTSTART:20260905T052500Z
DTEND:20260905T060000Z
DTSTAMP:20260911T054458Z
UID:session/Dvb1eRR7DUKKgJUE67SSHM@hasgeek.com
SEQUENCE:4
CATEGORIES:30 mins talk
CREATED:20260901T042321Z
DESCRIPTION:## Session description\n\nI keep getting asked some version of
  the same question - is this smaller\,\ncheaper model good enough to repla
 ce the hosted one - and I've had to answer it three times this year\, and 
 got it wrong on the first pass every time\, not because the model was bad 
 but because I was reading the wrong number. Tesseract\, a free local OCR e
 ngine\, looked almost as good as hosted frontier OCR APIs on word accuracy
 \, and I nearly shipped it\, until I checked the one thing the product act
 ually needed: could it still read question numbers\, the string that links
  a question to its answer. It recovered 68% of them on Hindi and 39% on En
 glish against 100% for every hosted model\, and the failure had a signatur
 e\, `*896.` coming back as `+896.` - same benchmark\, same model\, and the
  number everyone reports said close call while the number that mattered sa
 id not close at all. The same mistake showed up twice more this year\, onc
 e in an open embedding model I ended up choosing over a hosted one\, and o
 nce in a small vision model I added to a video pipeline and then had to ri
 p back out\, because the verifier built to catch its errors was catching a
  rule\, not a mistake. I'll walk through all three\, then run a two-minute
  script live on stage that reproduces the core failure - a page scoring 96
 %+ "accurate" while every ID number on it is wrong - so you can check whet
 her your own metric has the same blind spot.\n\n## Takeaways\n\n1. How to 
 pick the number that actually decides "good enough" before you run the ben
 chmark\, not after - because the default metric often isn't it\, and a sin
 gle pass/fail score lets the wrong model win.\n2. What it looks like when 
 a "small model checks the big model" setup goes wrong\, including a safety
  net whose biggest catch turned out to be a story it was built to tell\, n
 ot a thing it actually found.\n\n## Audience\n\nEngineers and technical fo
 lks who have to make this call and then defend it to someone else - a mana
 ger\, a client\, a teammate who disagrees. No ML research background neede
 d\; every method here is something you could run this afternoon.\n\n## Bio
 \nI'm Prudhvi Krovvidi\, a Data Scientist at Straive\, working at the\nint
 ersection of Data Science\, Generative AI\, and AI-driven applications. I\
 nenjoy turning data and emerging AI technologies into practical solutions\
 ,\nwith hands-on experience across LLMs\, RAG\, document intelligence\, co
 mputer\nvision\, and analytics.\n\nOutside of work\, I enjoy riding\, expl
 oring new places\, and experimenting\nwith new technologies. I'm naturally
  curious and enjoy learning by building\,\nbreaking things\, and figuring 
 out how they work. I'm also passionate about\nsharing what I learn and con
 necting with people from the wider data and\ntechnology community.\n\nLink
 edIn: linkedin.com/in/prudhvi-krovvidi\nEmail: kprudhvi71@gmail.com\n\n## 
 Draft slides\nhttps://prudhvi1709.github.io/talks/fifthel-05-09-2026/index
 .html
LAST-MODIFIED:20260901T043208Z
LOCATION:Hyderabad
ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com
URL:https://hasgeek.com/fifthelephant/agent-harnesses-small-models-hyd-mee
 tup/schedule/good-enough-for-what-Dvb1eRR7DUKKgJUE67SSHM
BEGIN:VALARM
ACTION:display
DESCRIPTION:Benchmarks that lie: when “good enough” models fail the th
 ing that matters in 5 minutes
TRIGGER:-PT5M
END:VALARM
END:VEVENT
BEGIN:VEVENT
SUMMARY:Break
DTSTART:20260905T060000Z
DTEND:20260905T062500Z
DTSTAMP:20260911T054458Z
UID:session/14ByWANRWaW3Ps853TCyrR@hasgeek.com
SEQUENCE:2
CREATED:20260901T042513Z
LAST-MODIFIED:20260901T043223Z
LOCATION:Hyderabad
ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com
BEGIN:VALARM
ACTION:display
DESCRIPTION:Break in 5 minutes
TRIGGER:-PT5M
END:VALARM
END:VEVENT
BEGIN:VEVENT
SUMMARY:Gantry: building agents that work like co-workers
DTSTART:20260905T062500Z
DTEND:20260905T070000Z
DTSTAMP:20260911T054458Z
UID:session/RrXYnf2VKyMAwAB7CjUv7k@hasgeek.com
SEQUENCE:5
CATEGORIES:30 mins talk
CREATED:20260901T042328Z
DESCRIPTION:Description\n\n  Every client project at KnackLabs wanted the 
 same thing: an agent that works like a coworker. It knows who it is talkin
 g to\, whether the message comes from Slack\, Teams\, or inside the produc
 t. It remembers past conversations with that person. It has the access an 
 employee would have\, and only that.\nSo it became Gantry: a self-hosted a
 gent runtime that teams embed through an SDK\, call over an API\, or reach
  from the chat tools they already use\, with voice planned next.\n\n  This
  talk goes through the harness pieces that held in production and the ones
  that turned out to be ceremony. Cases include an LLM permission\n  classi
 fier that cancelled a tool call at 08:53 and allowed the same call at 10:0
 1\, jobs that pause on a denial and resume after a crash\, and audit\n  re
 cords that still make sense weeks later. Each case covers what was tried\,
  how it failed\, and what replaced it.\n\n  Takeaways\n\n  Build the agent
  like you would onboard a coworker: one identity across every channel\, me
 mory that follows the person rather than the chat window\, and access gran
 ted up front rather than asked for at each step. Put all of that in the ru
 ntime\, not the prompt\, so it holds when nobody is watching.\n\n  Audienc
 e\n\n  Engineers and founders running agents in production or getting ther
 e. No ML background needed.\n\nBio\nRavi Kiran Vemula\, VP of Engineering 
 at KnackLabs. Builds Gantry\, an MIT-licensed agent runtime being prepared
  for open-source release\, and runs production agents on it for enterprise
  client work.\n\nPresentation\nhttps://claude.ai/code/artifact/5fcf9768-7e
 dd-4d69-9452-514b31eb5710
LAST-MODIFIED:20260901T043225Z
LOCATION:Hyderabad
ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com
URL:https://hasgeek.com/fifthelephant/agent-harnesses-small-models-hyd-mee
 tup/schedule/gantry-openclaw-for-your-company-RrXYnf2VKyMAwAB7CjUv7k
BEGIN:VALARM
ACTION:display
DESCRIPTION:Gantry: building agents that work like co-workers in 5 minutes
TRIGGER:-PT5M
END:VALARM
END:VEVENT
BEGIN:VEVENT
SUMMARY:Feedback and topics for the next meet-up
DTSTART:20260905T070000Z
DTEND:20260905T071500Z
DTSTAMP:20260911T054458Z
UID:session/8BrDmwZHx86aWca836c45f@hasgeek.com
SEQUENCE:3
CREATED:20260901T042546Z
LAST-MODIFIED:20260901T043240Z
LOCATION:Hyderabad
ORGANIZER;CN="The Fifth Elephant":MAILTO:no-reply@hasgeek.com
BEGIN:VALARM
ACTION:display
DESCRIPTION:Feedback and topics for the next meet-up in 5 minutes
TRIGGER:-PT5M
END:VALARM
END:VEVENT
END:VCALENDAR