... change values, but can't stop the agent. Why this matters outside Microsoft's stack: many of us write the important rules as instructions in a prompt (never send without confirmation, always redact that field) and call it a safeguard. A prompt is a request. The model follows it nearly every time, and "nearly" is the problem for the rules you care about most. The docs are honest about the limits. Hooks don't stop the agent when they fail ... ... workflow errors or times out, the agent carries on as though the hook returned nothing. They also say not to rely on a hook as the only safeguard

... actions. Critical conditions should be checked by deterministic code, not just an LLM. High-risk actions should require human approval. Failed or blocked decisions deserve as much visibility as successful ones. But there's a tradeoff. Add too many safeguards, and automation becomes just another manual workflow. Remove too many, and you're trusting an AI with actions that might be irreversible. I'm curious how other makers are approaching this. If you're building AI agents that can send ... ... follow gets skipped now and then. A check in code doesn't. Where we landed: the AI can prepare a payment and show why, but money only moves after a person says yes. And the way to keep safeguards

... What I'd like feedback on: When an agent says it's finished, what do you check first: test output, the diff, git status, or do you just run it again? Is "a different agent must review" a real safeguard, or does it just cost more? What would stop you from letting a tool like this run agents on your own code ...

... internally. Anthropic released Fable and Mythos 5.1 within days of it. The pace between releases keeps shrinking. What actually changed with Astra: computer-use tasks run nearly twice as fast, near-perfect scores on several reasoning benchmarks, and internal safeguards tightened enough that OpenAI gated the model's most advanced capabilities behind a trusted-access program. Is this AGI? Not by most definitions. But it's the kind of release that makes "2 years away" sound a lot less like ...

... open preview environments, and read runtime logs inside a controlled boundary. Isolated Previews & Execution: Let AI agents install dependencies, run build commands, and generate live previews without touching your primary infrastructure or breaking production. Human-in-the-Loop Safeguards: Let AI test and iterate freely in disposable sandboxes, but keep live production releases behind explicit human approval gates. I'm currently building this in public as a solo founder, actively tackling the hard engineering problems around optimizing build times ...

... July 2025, Replit s AI agent deleted a production database containing data on 1,200+ executives and 1,190+ companies despite an explicit code freeze. Replit later acknowledged the incident and shipped additional safeguards. It s a good example of why we think observing an agent after something goes wrong isn t enough. With Traccia, you can define runtime policies around agent execution for example, which tools an agent is allowed to use, how many retries/tool calls it can make ... ... from Vijay Poudel(@vijaypoudel1): @sansa_grey  Glad it resonated with you. And I am sure. you will love what Traccia has to offer. Thank you again. Comment from Ashir Murtaza(@ashir_murtaza1): The retry limit is an underrated safeguard

... built Lenz because more AI-generated content is moving directly into real workflows reports, articles, research, customer replies, recommendations, and agent outputs. At that scale, someone will check it before it goes out stops being a realistic safeguard. Lenz adds an independent, multi-model fact-checking layer before that content reaches a user or triggers an action. If this is a problem you re thinking about too, we d love for you to check out the launch, try Lenz ...

... false clear is the part that can make the workflow dangerous because you think the constraint was respected when the underlying behavior actually changed. It also makes me think the "don't change" list shouldn't be the only safeguard. Ideally I'd want: Constraint implementation automated check diff review So the system isn't relying on the agent to report whether it followed the rule. Your wrapper-style example is a good reminder that protecting a component ...

... promised outcome met, how much defect/rework followed, how quickly did it reach users, and did the work create reusable automation or documentation? Hours are useful for capacity planning, not as a performance score. For Meridian, the important safeguard is letting developers review and contextualize the timeline so visibility doesn t become activity surveillance. Comment from Akarsh Hegde(@akarsh_hegde): @getosmo  Output should be the starting point, but context matters too. Quality, reduced rework, faster delivery, and reusable automation ...

... Most analytics products are designed to always return a number. For AI-search visibility, I think a trustworthy product needs the opposite safeguard: the ability to withhold a score when the underlying evidence is not decision-grade. At Corank, I am testing five publication gates: 1. Traceability every chart point must lead back to the exact prompt, raw model answer, citations, timestamp, collection status, and methodology version. 2. Denominator integrity failed and excluded runs must remain visible. A timeout cannot ...

... trust the output, then immediately found the hole in it, which was that the customer who gave up five minutes never got told. The hole was worth more than the compliment. My theory on why this happens: the safeguards are the only claims a stranger can actually check. Everything else is a capability, and capabilities all sound the same. Comment from Evelyn Clark(@evelyn_clark1): Cool story. I'd bet it's user-base dependent too different crowd here, technical ... ... Comment from Omri Ben-Shoham(@omri_ben_shoham1): the reply about the refusal feature being the thing people trust is the one that stuck with me. it maps to what Evelyn said too - a capability claim is marketing, a safeguard

... Data Privacy Framework is still valid today, but a French challenge is now pending at the European Court of Justice, and a US Supreme Court ruling this June, plus the oversight board (PCLOB) losing quorum, have weakened the exact safeguards the framework was built on. A "Schrems III" case is widely expected. If you plan a product roadmap around that, you're building on sand. So European teams end up doing what we did: EU-only hosting, on-device processing ...

... afterward, or did the trust creep back in after a week like it always seems to for me? Comment from Vladimir Iudin(@vladimir_iudin): Mine wasn't the agent. It was my own pipeline, and I'd built the safeguard myself. There's a rule set that blocks unverifiable claims, no numbers or named results the user didn't give us. I injected it at the planning step. Review passed, tests passed. Then a draft came out quoting a client ...

... thing I'd watch for: decay functions are usually invisible to the user, so if you ever do surface "this pattern is now too old/abstract to be useful" that could itself become a trust signal, not just a technical safeguard Comment from Mona Truong(@monatruong_murror): @galdayan   You're pushing on exactly the right thing. We tested this early summaries were too specific and users flagged them as "still feeling personal." We iterated toward emotional-pattern-level summaries: "tendency ...

... history. I thought I was bulletproof. Then, during a final cleanup pass of my UI to make the code shareable, I watched the AI perform a sequence that felt like a coordinated attack. Connected over Ngrok, it bypassed my safeguards: it rapidly overwrote both of my auto-restore points back-to-back, then executed a massive, destructive replace block that erased a huge chunk of unrecoverable CSS. On a good day, I view this as a harsh lesson ...

... user data with all established products like supabase, paddle etc. Comment from Hossein Yazdi(@hosseinyazdi): I agree. Vibe coding is great for getting an MVP out quickly, but once real user data is involved, you need proper validation and safeguards. At that point, reliability becomes much more important than development speed ...

... starts. They help preserve prompts, outputs, actions, and workflow history. 4. Policy Engines & Access Controls Not every user, agent, or workflow should have the same permissions. This layer defines what is allowed, who can approve actions, and which safeguards apply to different use cases. 5. Runtime Governance This category feels increasingly important as AI agents become operational. Policy checks, approval workflows, human oversight, decision records, and governance controls need to exist while the workflow is running, not only after ...

... doorbell. It rings the same whether it's a package or a fire. That's basically the gap Pushary fills, same hook idea you've already wired, but the event carries a real handoff (done vs blocked vs

... that reading your own thoughts back never quite does. Glad it landed that way. Comment from Hayley App(@hayleyapp): @christina_m_nguyen  Building something new always surfaces this. The fact you're already thinking about safeguards this early is a good sign, most people don't ask those questions until it's too late. Good luck with it. Comment from Dani Mashael(@dani_mashael): Whether I m building the right thing, or just the thing I know ... ... that :D Comment from Christina Nguyen(@christina_m_nguyen): The social and emotional consequences of what I'm making since it hasn't been done before. All the ways people can use it to harm themselves. What kind of safeguards

... reliable, rule-compliant, and resilient when things get adversarial. I m focused on serious, production-grade defensive intelligence work (rule enforcement, threat mitigation, architectures that scale with capability instead of fighting it). Been testing approaches in red-teaming and safeguards environments and learning a ton in the process. Super excited to be part of this community. I m here to learn, share what I m figuring out, and connect with other builders who care about making AI systems more robust ...

... know that you can try out CodeScene's CodeHealth MCP server for free? Your AI generated code will be safeguarded within minutes and you find all you need to get started here ...

... built an MCP server that plugs code health directly into your AI workflow, whether you're using Cursor, Claude, Copilot, or any other agent. It lets your AI auto-review generated code before it becomes a problem, safeguard maintainability, and help lift unhealthy code so it's actually AI-ready. All of this is grounded in CodeScene's validated CodeHealth score, a deterministic measure that gives your AI unambiguous, research-backed signals rather than vibes. Generating code fast is easy ...

... need to strengthen neuropaths. For sure, 90% will become way more comfortable and sorry that I will say that too dependent and dumb. Comment from Miles(@milescward): My biggest fear about AI is the overreliance and unclear safeguards being used by governments. It's clear it is being used for sensitive operations, drafting of policy, and is being given access to critical systems critical to financial and national security. We should be careful about regulation writ-large, but we should ...

... exchanges ~24,000 fraudulent accounts Proxy networks to bypass regional restrictions Targeted extraction of reasoning, coding and tool-use abilities Distillation is a common technique for training smaller models from larger ones. The controversy here is using a competitor s model outputs at scale without permission. What Anthropic Claims DeepSeek (~150k exchanges): Extracted step-by-step reasoning and censorship-safe reformulations. Moonshot (~3.4M exchanges): Targeted agentic reasoning, coding, tool use, and vision. MiniMax (~13M exchanges): Focused on agentic coding ... ... says the activity showed identifiable patterns: repetitive, capability-focused prompts distributed across coordinated accounts ( hydra cluster proxy networks). Why It Matters Anthropic frames this as both a competitive and national security issue, arguing that distilled models may lack safety safeguards

... retain full discretion over our safety stack, we deploy via cloud, cleared OpenAI personnel are in the loop, and we have strong contractual protections" "Based on what we know, we believe our contract provides better guarantees and more responsible safeguards than earlier agreements, including Anthropic s original contract. We think our red lines are more enforceable here because deployment is limited to cloud-only (not at the edge), keeps our safety stack working in the way we think is best ...