AI meetup (Boston): GenAI, LLMs and Agents — AICamp
Table of Contents
1. Logistics
- Event
- AICamp Boston — GenAI, LLMs and Agents
- When
- Thursday, 2026-07-23, 17:30–20:00 EDT (mixer after)
- Where
- Microsoft NERD, 1 Memorial Drive, Cambridge, MA 02142
- Room
- Deborah Sampson & Thomas Paul
- RSVPs
- 121
- Partners
- Microsoft
- Details
- aicamp.ai event page
- Discord
- AICamp community
2. Related notes crosslink
Same venue + same night, different event:
- Boston Generative AI Meetup — The Harness for AI Agents (2026-07-23) — the other NERD meetup tonight; harness-first framing (panel + Jesse Waites tech talk).
Recent predecessor:
- AI Tinkerers Boston — GTM Agentic AI Launch (2026-06-29) — "harness, not model"; durable execution, secret-brokering, harness-eval attribution teardown.
Research to align against, by this event's topics:
- Fleets / scale
- agentic-2026.
- Provenance / lineage
- containment-mapping · declared-vs-effective-state · agent identity attestation.
- Agent memory
- 2026-agent-memory-systems · agent-memory-institutional-knowledge.
- MCP / tools
- code-graph MCP survey.
- Eval / observability
- terminal-bench.
3. Agenda
| Time | Item |
|---|---|
| 17:30–18:20 | Check-in, food, networking |
| 18:20–18:30 | Welcome / community update |
| 18:30–20:00 | Tech talks + Q&A |
| 20:00 | Open discussion & mixer |
4. Talks
4.1. Building Agent Systems At Scale with VertexAI and Gemini people
Thesis: 1 agent = demo; 100 in prod = systems problem. Self-scaling fleet, failure recovery, agents checking each other's work; patterns on Gemini + VertexAI.
- Watch for
- what "check each other's work" means as a contract (voting? adjudication? blast radius on a bad agent). Recovery = retry vs compensating action.
- Align
- AI Tinkerers — Poliakov (durable execution / resume-on-crash); the governance-tuple framing (agent / reviewer / arbiter).
[ ](no term)
4.2. Can We Rebuild AI's Source of Truth? people
Thesis: AI artifacts (code, data, weights, packages, infra) lose provenance unlike software. Two systems primitives make lineage automatic + objective; runtime observation captures provenance w/o code changes or infra lock-in.
- Watch for
- the two primitives; how runtime-observed provenance stays sound (tamper model, completeness vs sampled capture).
- Refutation
- "no code changes" — what escapes observation? out-of-band data mutation, side channels.
- Align
- containment-mapping and declared-vs-effective-state — runtime-observed vs declared state is the same lineage problem.
[ ](no term)
4.3. How AI Agent Memory Works people
- Watch for
- memory taxonomy (working / episodic / semantic), write vs retrieval policy, eviction, staleness/consistency guarantees.
- Align
- 2026-agent-memory-systems · agent-memory-institutional-knowledge; AI Tinkerers — Zhu (markdown memory → RDBMS, compaction/cache of context).
[ ](no term)
6. Claims to test refutation
- "Agents checking each other's work" improves reliability (Lust)
- a bad agent whose error is ratified by its checkers (correlated failure), or a checker that adds latency/cost without changing an outcome.
- Runtime-observed provenance needs no code changes (Geyer)
- a mutation applied out-of-band (side channel, manual edit) that the observer never sees.
- Provenance capture is complete, not sampled (Geyer)
- a lineage gap where a sampled/dropped event breaks the chain.
- (no term)
7. Open questions / Q&A
- Cross-checking: is it voting, adjudication, or a single arbiter — and what is the blast radius when the checker is the one that's wrong?
- Recovery: retry the failed step, or a compensating action? Idempotent?
- Provenance tamper model: who can forge or drop a lineage event, and is the chain hash-linked (tamper-evident) or just logged?
- Memory: write-vs-retrieval policy, eviction, and staleness — what consistency does a multi-agent fleet get on shared memory?
8. Contacts to follow up network
| Name | Org | Contact | Thread |
|---|---|---|---|
9. Actions
[ ]
10. Scratch