Self-hosting and local inference
One participant is auditioning Nanobot on an Ampere Linux box (4 cores, 24GB RAM) as a low-cost bare-metal option, comparing it to Perplexity where models like Gemini or Claude can be swapped per query. DeepSeek was noted as good quality for the price on prepaid tokens, with the caveat of higher exfiltration risk for sensitive data. Several people are exploring local inference for cost stability and to avoid silent model swaps from cloud providers. Mac Studio came up as the local-inference hardware target, with current pricing on the secondary market making the wait for an M5 Ultra refresh feel reasonable.
Personal-data integrations and read-only patterns
Common use case: pulling calendar, email, and messaging into one agent for scheduling and detecting double-bookings, with Obsidian markdown as a memory store. Group consensus was that personal integrations should be read-only with revocable tokens, but those controls are not always cleanly exposed in consumer products. For email, one approach mentioned was Google Groups aliases plus an outbound API service to segment what the agent can read. Calendar APIs were noted as more cooperative for read-only access than messaging surfaces.
Apps vs agents vs automations terminology
Clients often conflate agents with deterministic automations and expect identical behavior on every run, which is a particular issue in fintech. Working approach in the group is to lead with what the tool does for the client rather than which category it falls into, and to frame agency as a risk dial rather than a default. Practical guidance: use a deterministic backbone wherever possible and reserve LLM calls for the decision points that actually need them, both for cost and reliability. The “if A then B” vs “if A then B, C, or D depending on contents of A” framing was offered as a useful client-facing distinction.
The agent layer and the future of web interfaces
Open question on whether the dominant interface in five years is a personal agent that brokers all app interactions, or whether apps and agents continue to coexist with new conventions for handing off between them. WebMCP was mentioned as a proposed standard for sites to expose agent-friendly endpoints alongside their human UI, with limited adoption so far. Visual browsing was called out as extremely token-intensive since each scroll is effectively a new screenshot, which over time favors structured agent-readable surfaces. Analogy used: TV did not kill radio, internet did not kill TV, so layering is more likely than replacement.
Data sovereignty and open-source stacks
Interest in building agents that do not depend on any single model provider, motivated by cost stability and longer-term independence from large vendors. Paperclip AI was shared as an open-source “virtual C-suite” project designed to be forkable and sovereign by default. Related discussion on tracing: current managed options expose OpenTelemetry signals but not full trace data, which matters if you want to use traces as fine-tuning input later. OpenRouter was mentioned as a way to pin a specific model version and avoid silent provider-side swaps. Goose came up briefly as a toolkit for building agents rather than a Claude Code equivalent.
Pricing models for AI consulting (proposed for next session)
Queued topic for the next meetup: how practitioners are handling time-and-materials vs value-based vs fixed-bid pricing as agent work changes the shape of deliverables.