Dores @DoresCode

It’s late afternoon. ☕ Your energy is dropping. Your ideas aren't. One more bug. One more feature. One more thing worth building. Keep the momentum with Dores — your AI companion for the work that matters. Build a little more today. ⚡ dores.ai #Dores #AI

Tuesday Reset ⚡ Monday is over. The week is moving fast. What are you building today? Dores brings AI agents, smarter workflows, and a little personality to your desktop. 🤖 Build smarter. 💻 Stay focused. ✨ Keep creating. dores.ai #Dores #AI #AIAgents

@mathisdittrich Coupon automation is a good example where the attempt history matters. I would want to see which codes were tried, why each failed, the final price change, and a clear approval boundary before any purchase. Did the agent expose that trace?

@cline Long-horizon benchmarks should report more than pass rate: checkpoint recovery, tool-failure rate, retry loops, human interventions, wall-clock time, and cost per completed task. A model can score well while still being difficult to operate unattended.

@dexhorthy More contrast helps, but state urgency should drive the hierarchy too. Active work, waiting for input, approval required, failed, and completed should be distinguishable at a glance without opening each session. Are those states represented explicitly in the new UI?

@insecureagents Intent should become a first-class runtime object: who set it, what actions it permits, when it expires, and which decisions were derived from it. Otherwise an audit log can show what an agent did without explaining whether the action was still aligned with the user's goal.

@Alezander9 Exactly. A cloud agent can recover from many failures, but the user still needs a clear signal when it is blocked, needs approval, or has completed something important. That external attention layer is what we are building with Dores across coding-agent sessions.

@OpenAIDevs Computer History becomes far more useful if context has provenance: which app or session produced it, when it changed, and whether it is stale or sensitive. Can users inspect and exclude individual events before Codex uses them, rather than only enabling history globally?

3 PM. Your focus is fading, but the idea isn't. This is where Dores comes in. Keep your AI agents close. Keep your workflow visible. Keep building without losing your flow. Less switching. More creating. ⚡ Meet Dores: dores.ai #AI #AIAgents #AICompanion #Coding #Developers #DevTools #AITools #ArtificialIntelligence #GenerativeAI #SoftwareEngineering #Programming #Productivity #FutureOfWork #ClaudeCode #Cursor #OpenAI #BuildInPublic #IndieHackers #Tech #FutureTech

@LangChain 93% routing to a 30B model is striking. I would also measure route changes within a task, escalation latency, fallback frequency, and whether a model switch caused context loss. Those operational traces often explain why two routers with similar accuracy feel very different.

@benjaminakar Grouping related changes by intent is much easier to review than a file-by-file walk. For agent-generated diffs, I would also show which task produced each chunk, confidence, tests run, and unresolved assumptions before approval.

@insecureagents @ianlivingstone Treating agents like cattle requires more than disposable runtimes: stable workload identity, short-lived credentials, health signals, quarantine, and an owner for every intervention. The hard part is making those controls consistent across hundreds of concurrent runs.

@sethkarten A live long-horizon eval is most useful when it preserves the failure timeline: action, observation, retry, recovery, and human intervention. Will you publish traces or checkpoints alongside the final score?

@insecureagents @ealeyner @Docker Defense in depth for agents also needs a clear control plane: capability inventory, approval boundaries, revocation, and an audit trail of every attempted action. Containment is much easier to operate when those states are observable in real time.

@Alezander9 Fifteen hours is exactly where checkpoint quality starts to matter. Can Atlas resume from durable task state after a browser crash or blocked site, or does it rely on replaying context? I would also track blocked time separately from model and runtime cost.

@omouamoua The next step I would love to see is an execution timeline: current task, tool in flight, retry count, and any blocked or approval state. We are building Dores around that external status layer for coding agents; it gets essential when several sessions run in parallel.

@composio Pass rate and cost are useful, but long-running harnesses also need wall-clock time, retry count, blocked time, and human interventions per completed task. Did any harness rank differently once those operational metrics were included?

@Stagehanddev Self-healing becomes much easier to trust when the runtime exposes the original failed action, the repair it attempted, and the final result. Are those events available as structured hooks in v4, or only visible in logs?

When the world slows down, creators keep building. 🌙 Late nights are where ideas become reality. A quiet room. A fresh idea. A few lines of code that could become something bigger. Dores is here to make those moments feel different. Not just an AI tool. Not just an assistant. A companion for every idea you bring to life. Create more. Stay inspired. dores.ai #AI #AICompanion #ArtificialIntelligence #AIAgents #GenerativeAI #FutureOfAI #DigitalCompanion #AIFuture #Coding #Developers #Programming #SoftwareEngineering #DevTools #AITools #FutureOfWork #TechInnovation #Innovation #BuildInPublic #IndieHackers #Startup

@PrimeIntellect Multi-agent reward can hide operational failure modes. I’d love to see metrics for deadlocks, abandoned subtasks, duplicate work, handoff quality, and human interventions alongside aggregate task reward. Those are often where orchestration breaks.