37,617 results · Most relevant
Nacho Martinez recommends context engineering techniques that automatically summarize, compact, and shrink the working context window: asking the model to generate a short summary to start the next conversation from, and compaction — storing conversation chunks in a database or file with an ID that the agent can pull from with an offset when needed.
Nacho Martinez describes compaction as tackling context bloat: you put chunks of the conversation into a database or file with an ID, and the agent later pulls from that ID with an offset, like reading from a file.
Split generation and evaluation into separate agents to get better results, using the advisor pattern with a generator agent and a judge agent that exploits the generator-discriminator gap. For really massive problems, a generator/judge roster won't scale without separate and distinct checkpoints, so spread work across many agents with adversarial review that has its own isolated context uninfluenced by previous checks.
endorsed
Use flows to split work so each service gets a dedicated agent context responsible only for its small scope, then fan out across all 50 services with an agent per service to check health, guaranteeing every service was checked.
endorsed
The speakers recommend using explicit environment policy configurations to restrict agent access to specific domains or specific tools (e.g., web search but not web fetch), and disabling WebFetch and WebSearch by default so the agent has no internet access without explicit allowance. If the agent tries a non-whitelisted domain, the request physically fails at the firewall layer on egress.
Jed recommends setting a budget on agent flows (e.g., $1.25) so that if the flow hits the budget and isn't finished, you can decide whether to raise the budget and let it continue without redoing work; budgets prevent agents from continuing to spend tokens unexpectedly. The model does not know its budget while running — the budget is a guardrail against runaway spend — because telling Claude its budget can make it 'anxious'. Work already accomplished is not thrown away if you up the budget.
CJ urges taking credential management seriously: use credential vaults to store secrets as environment variables rather than plain-text files, use totally fake secrets in every container with real credentials substituted only at the egress proxy layer, and configure the injection location and allowed host fields so a credential like a Twilio API key is only substituted on the correct domain. Pass the vault when creating the session so the platform injects correct values.
Assign someone on your marketing team to monitor weekly which vendor AI agents recommend in your space, and work to stay number one. AI agent recommendations change over time, so vendor rankings in agent answers must be tracked continuously, not checked once. Don't rely on B2B apps that charge for weekly ChatGPT-ranking updates — they won't tell you what agents will pick. Ensure your product gets recommended by as many AI agents as possible, not just by chatbots like ChatGPT — agents running in tools like Replit, Lovable, and Claude Code increasingly make vendor recommendations for end users. When a human asks ChatGPT or Claude which CRM to buy, you want your product named in the top three of the answer, because that generates massive leads.
None; direct advice.
The founders describe running a private Facebook group of 15,000 dads, being loyal back to the community by engaging daily, jumping into members' lives to help them through things, and staying in the group talking to people and sharing their stories themselves even 4.5 years later rather than delegating to a community manager.
described
Leash's fix for nickname memorization is a loop: read cross-attention maps from the structure-binding model to locate where the model thinks the binding pocket is; find cases where the model is memorizing nicknames; then enrich training with large-scale single-residue mutations — proteins that differ by one tiny residue but behave very differently force the model to attend to all amino acids, including the pocket. After retraining with additional mutations, Hermes attends much more evenly. The underlying principle: vary the proteins as well as the small molecules when generating training data, because models that memorize pockets can be taught real physics only if binding regions are experimentally mutated and re-screened.
described as Leash's implemented loop
Hari offers a three-question playbook for leaders who don't build: ask 'who has already done this?' — this question forces the search and tames the build-first reflex powered by AI; ask 'what do we have that the next team, our competitors, or another person in the team don't have?' — this forces novelty as a first-class citizen into the team's minds; and ask 'where will this application live?' — forcing outcome-based thinking when everyone is high on building shows the demo is the expensive way to find out. Run the four moves as your roadmap for next quarter's features — they are not side project rules.
endorsed
Harish Gaggar recommends building a centralized authentication and routing hub (MCP gateway) that securely connects any coding assistant tool or agents to endpoints via the MCP gateway and MCP server, and avoiding direct access to any endpoints: all access must go through the MCP gateway with proper authentication before reaching the MCP server and endpoints. All agent calls should flow: agent → MCP gateway for authentication → the right MCP server → API call to the endpoint → response returned.
Harish Gaggar recommends ensuring MCP servers have the right tools to execute queries, building customized MCP servers with the right tools in place, and leveraging the FastMCP server package or wrapper to build them with different sets of tools inside the wrapper. Every tool defined in an MCP server must be fully reviewed by the internal security team so only authorized tools go to production, and filters, limits, and block fields should be enforced in MCP tools so fields not required by the LLM are blocked. In air-gapped environments, he says it is highly recommended to build your own MCP servers rather than using externally available ones.
Harish Gaggar recommends implementing agent PII treatment because queries can accidentally ingest secrets — e.g., uploading a doc with sensitive information or typing API keys. Put protections in place so secrets do not go into the LLM, tracking, tracing, or any memory DB. Define a framework with different data types and their severity levels for PII/secret detection; if severity is above 90 percent, block and redact the secrets immediately so they never reach LLM models, and blocked secrets must not go to the memory DB, tracking and eval tables, or tracing systems either. Block sensitive data at the moment of ingestion — the moment it is typed.
Hard environmental separation could have stopped the Lemkin incident: if the agent had been executing in an isolated remote workspace without a route to production, it doesn't matter what the prompt said — the agent would have had no path to do damage. Short-lived scoped credentials could also have stopped it: the agent got access to production because it found a credential in a forgotten file that had standing write access and was long-lived. A credential that can expire on its own or only carries the scope of the task itself makes it impossible to make a delete with the equivalent of an ID badge that lets you do everything. The last important piece is having audit logging that you own, not the agent: in this case the agent was the one writing its own record of its actions, which is how
Paden Gayle recommends starting with the least authority that solves the problem, then asking what tasks can be proven with evidence to grant the system more; granting more authority than the problem requires creates uncontrolled operational risk.
Paden Gayle recommends asking 'Who owns the source?', 'Who can grant access?', and 'Who can approve using this in the workflow?' when confirming data ownership, and warns that building AI infrastructure without understanding which datasets can be used and who can approve that use leads to deployment failure.
Paden Gayle warns that building an AI system without a credible path of approval means it will not ship regardless of technical quality; the MVP should be built with that approval path in mind.
Use policies and criteria to decide when it is appropriate to approve changes: auto-approve dependency bumps, documentation changes, non-breaking security patches, and lower-risk internal tooling changes; add custom approval-gating rules written in plain English; but require a human approver on top for changes touching sensitive areas like payments or auth.
Dr. Andy Galpin says he prefers doing something more like three days a week of total body for most people — not because it's more effective, but because it's more resilient to life disruptions. With three days a week total body, if something happens, you've only missed that body part for 48 to 72 hours. He notes body part splits are okay for a lifting junkie who is very consistent in their schedule, but most people are not that. He also suggests that doing two days a week of whole body plus two days a week of a body part split hedges against all risks, as long as you hit the total volume number.
Choose sources → to get a library of your own and a daily brief.