Barak Fargoun

CTO @ Foundational · — enterprise context for AI that has to know what it’s touching

[Barak Fargoun @BarakFargoun]

@bygregorr @cyrilXBT nasty failure mode. agent can look internally coherent while missing the one field that changes the answer. context completeness needs a first-class signal.

[Barak Fargoun @BarakFargoun]

@MartinSzreter @CLEMENTAIEXPERT green pipeline can still publish wrong semantics. schema/contract change should surface affected downstream assets before consumers find out.

[Barak Fargoun @BarakFargoun]

@kidehen context without semantics stays ambiguous. and semantics alone go stale without live lineage. useful layer ties meaning to current data and impact.

[Barak Fargoun @BarakFargoun]

@KPeyanski stale vs bad data. that's the tradeoff. rejecting an impossible meter jump helps, but downstream still needs to know the value is old, not current. is freshness surfaced beside the reading?

[Barak Fargoun @BarakFargoun]

@ayshriv kafka events as APIs. catching a breaking field change before five consumers find it in prod is the hard part. where does that check live for you today?

[Barak Fargoun @BarakFargoun]

@rodydavis @typesafeai keep candidate selection close to the database. fewer handoffs means fewer chances for retrieval and reranking to disagree on meaning.

[Barak Fargoun @BarakFargoun]

@Graphed approve/decline step. if the source call never becomes structured CRM state, enrichment is just polishing a hole in the capture path.

[Barak Fargoun @BarakFargoun]

@xevrion_the1 senior skill is the verification path, not the prompt. confident answer with no source context and no way to inspect the underlying data is still a production bug.

[Barak Fargoun @BarakFargoun]

@murugavell “data before models” is still the unglamorous bottleneck. one operational check: which input rule failed, on which slice, and what changed since yesterday? otherwise garbage in stays invisible. what did Simons measure first?

[Barak Fargoun @BarakFargoun]

@alveejack1 “yesterday’s environment” becoming today’s wrong world is a good freshness test. every capture should carry where/when plus an expiry, not just a new snapshot. who owns that freshness budget?

[Barak Fargoun @BarakFargoun]

@RaulJuncoV read-only graph view over relational data feels like the pragmatic middle. same source of truth is the payoff. how do you expose query cost or stale relationship state before teams rebuild a sync pipeline?

[Barak Fargoun @BarakFargoun]

@marquisehurtt same-agent data is the tell. demo looks great when the generator already knows the path. what independent fixture do you trust before you widen production blast radius?

[Barak Fargoun @BarakFargoun]

@xqliu timezone normalization is the edge. types and tests protect shape, not intent. what contract check catches a valid schema change that is semantically wrong?

[Barak Fargoun @BarakFargoun]

@suraj_sharma14 “you still own the blast radius.” a tool call that writes files can pass retrieval checks and still be wrong. do you gate review on side effects too?

[Barak Fargoun @BarakFargoun]

@jacob_luetzow 181,163 test lines next to 140,655 app lines. serious net. but what catches a missing code edge when a green suite never covered that path?

[Barak Fargoun @BarakFargoun]

@samuelstroschei “company brain” only gets real at the data-source joins. how are you catching stale context before an agent turns an old source into a confident answer?

[Barak Fargoun @BarakFargoun]

@abhijitghoshin 99.9% pass rate on the checks you wrote. regulator incidents came from fields no rule covered. rule coverage beats a green score.

[Barak Fargoun @BarakFargoun]

@LimestoneHQ dashboards answer questions someone already measured. she answered the ones nobody anticipated. agents stuck on the dashboard miss the ad-hoc query path that runs the deal.

[Barak Fargoun @BarakFargoun]

@DSPyOSS asking a frontier agent is like managing a person. hand-writing Jev for every branch is assembly. what lasts is the policy between them: what gets decided, logged, reused.

[Barak Fargoun @BarakFargoun]

@johnmorganFL model is only one layer. how far an exploit travels is permissions + egress + monitoring. tool scope is the blast radius.