35,684 results · Most relevant
A method for debugging production agents: define failure-mode categories captured in order of occurrence, filter traces by score conditions (e.g., dispute not resolved, security risk flags) instead of scanning hundreds of thousands of traces, and save filtered views for reuse.
Agents need a way to wake up in response to events. Subscriptions is a primitive added across their agent harnesses. They have a bot that watches a dashboard, and whenever a shim stops firing, that agent puts up a PR saying the shim can be cleaned up.
When a former customer stakeholder joins an account belonging to a peer, send that peer a signal — described as a great way to build morale and camaraderie inside the organization.
The customer and lost-deal job change searches are especially valuable for frontline sales leaders: consider the combined territory size of your five-to-nine direct reports and track job movements over time to see who is moving where and help your AEs prospect directly.
When a former colleague becomes a first-degree connection at a company you want to sell to, ask them to help you break into their org, even if you don't sell into their department. Sam McKenna frames this as an obvious move: 'you'd kind of be a fool not to' ask.
When a job change alert fires, leverage referrals and common connections to figure out your shortest path to that executive.
Save power users — super fans — to a list and reach out to them when they change jobs.
Set up the profile views search deliberately excluding seniority, geography, and other filters — any interest from anyone in your territory matters, regardless of seniority or role.
Never reference the profile view directly when reaching out — don't say 'hey, saw you viewed my profile.' Don't assume why a stakeholder viewed your profile either: they may be in another division, on another team, or caught wind of leadership discussing you; the message works either way.
Pamela Fox recommends giving LLMs a tool to do date math instead of expecting them to compute dates in their head, especially on leap years where date math is especially hard.
Pamela Fox recommends treating the JSON schema function definition sent to the LLM as prompt engineering — decide what types, descriptions, and parameter names will be best for the LLM. Include a name, description, and typed parameters in your JSON schema function definition.
Pamela Fox recommends using observability providers' built-in eval frameworks such as Langfuse, LangSmith, or Arize, since most observability providers have evals built in.
Pamela Fox recommends testing evals across different languages, whichever languages you're going to support, and including edge cases like ambiguous durations and slangy phrasing in eval test cases. She also suggests testing on leap years because date math is especially hard then, and including ambiguous user messages as test cases.
Pamela Fox recommends structuring eval test cases with a user message and checks that do string and list comparison, and testing a single tool call by checking whether it called the right tool and used the right arguments via string comparison.
To test tool-calling efficiency, give the model a word problem requiring multiple calculations (e.g., buying three shirts with tax), use a calculate function that takes a mathematical expression and returns the output, and add the system prompt 'do not do math in your head' — needed especially for GPT-5.5, which prefers doing math in its head.
Ask the buyer whether the spend decision is within their authority or requires a majority vote based on spend volume.
When a buyer says 'we don't have budget allocated', treat it as a value-proofing problem, not a dead end — budget often isn't allocated for anything new. Partner and multi-thread to pull from different budgets — e.g., pitch a talent-attraction use case to HR or the chief people officer.
When a champion says they'll present a recommendation instead of options, ask them the same question you'd ask yourself: why won't this deal get done? Ask the champion to anticipate objections, pushback, and existing vendor relationships that could sway the decision, then ask what you can do to prepare them.
Instead of asking a champion the open question 'what resistance might there be?', offer specific common resistances as options — it's hard for them to search every permutation, but they can pick from suggestions and it jogs their thinking. Sam adds that suggestive capability should be part of every stage of the sales process, because asking only open questions gets you only what's already in the buyer's mind. Maintain a running list of your deal-loss reasons, then filter that list by vertical, buyer, and deal size to suggest likely objections to your champion.
Freeze the state right before a PR was solved in an external repo and let the system run; if it resolves the task, you gain confidence the improvement came from model weights rather than future information leaking in.
Choose sources → to get a library of your own and a daily brief.