U G Murthy

Day 104: #AdaptiveAgent #BuildinPublic #AI The final result in markdown : ugmurthy.github.io/ai_agent_loops… Using the context-ref feature of adaptive-agent the markdown was converted to a html report ugmurthy.github.io/ai_agent_loops…

Day 104: #AdaptiveAgent #BuildinPublic #AI 4/5 : Using trace-session to trace the workflow youtu.be/znlbyQrGeoc

Day 104: #AdaptiveAgent #BuildinPublic #AI 3/5 The workflow produced both a research report and a paper-style HTML artifact, with a separate trace video showing how the session unfolded. youtu.be/C7sWol9qdFg

Day 104: #AdaptiveAgent #BuildinPublic #AI 2/5 The goal was not to hype agent loops. It was to stress-test the idea: what is actually new, what is a repackaging of older automation patterns, and where the architecture creates real leverage.

Day 104: #AdaptiveAgent #BuildinPublic #AI 1/5 Are AI agent loops just old wine in a new bottle? I used Adaptive-agent with a custom storm-researcher agent to critically examine the question using Stanford's STORM method.

Day 103: When agents delegate work, traceability has to cross the handoff. AdaptiveAgent traces can show parent runs, delegate names, child runs, child status, lease signals, last events, and failure reasons. #AI #AdaptiveAgent #BuildInPublic

@with_isabell Considering that the CLI and trace-session is shipped together anyone having access to CLI, has access to traces.

@murthyug "Traceability is also a safety feature" — this. The moment an assistant can spend money or act for someone, the trace is the trust. We made budget-exhausted and approval-pending first-class states too, not buried logs. Do you expose those to end users or keep them dev-only?

Day 102: Traceability is also a safety feature. AdaptiveAgent can surface approval signals, rejected tool calls, budget-exhausted tool calls, and runtime-injected policy or budget messages. checkout :ugmurthy.github.io/adaptiveagent.… #AI #AdaptiveAgent #BuildInPublic

@futureefoundry Check this out #1" target="_blank" rel="nofollow noopener">ugmurthy.github.io/adaptive-agent… - a 2 minute pitch or this to get started : ugmurthy.github.io/adaptiveagent.…

@murthyug Would love to feature this. Drop the link? 🚀

Day 101: #AdaptiveAgent #BuildInPublic #AI Implemented --context-ref a way to reference prior runs results and re-harvesting for followup tasks. This also sets up adaptiveAgent to implement memory. to know more about adaptiveAgent : ugmurthy.github.io/adaptiveagent.…

Day 103: Sometimes you need to inspect the context, not just the outcome. `trace-session --messages` can show snapshot-backed system, user, assistant, and tool messages. `--system-only` narrows the policy view. checkout :ugmurthy.github.io/adaptiveagent.… #AI #AdaptiveAgent #BuildInPublic

Just shipped my project for the @meshapi_ai Hackathon 🛠️ Built on Mesh API. #MeshHackathon hack.meshapi.ai

Day 101: Slow agent run? Trace the split: model duration, tool duration, snapshot save time, adapter latency, retry delay, request bytes, response bytes, and slowest tool spans. checkout :ugmurthy.github.io/adaptiveagent.… #AI #AdaptiveAgent #BuildInPublic

Just shipped my project for the @meshapi_ai Hackathon 🛠️ Built on Mesh API. #MeshHackathon hack.meshapi.ai

@ampcode - Great UI , nice combinations of models - what makes 'ultra' ultra when compared to 'high' - considering 'high' uses GPT 5.5 xhigh whereas 'ultra' uses GPT 5.5 high

Day 100: Agent cost needs attribution. AdaptiveAgent traces can break usage down by root run, provider, model, tool-output usage, tokens, reasoning tokens, and estimated cost. checkout :ugmurthy.github.io/adaptiveagent.… #AI #AdaptiveAgent #BuildInPublic

Day 99: When an agent fails, the first question is usually: where? `trace-session --view timeline` shows tool calls, inputs, outputs, durations, outcomes, child runs, and errors in execution order. #AI #AdaptiveAgent #BuildInPublic checkout :ugmurthy.github.io/adaptiveagent.…

Here is one example of how it helps x.com/murthyug/statu…

Day 98: Agent workflows are chains of decisions. AdaptiveAgent traceability helps answer: what happened, which tool ran, what failed, what it cost, and why the workflow ended where it did. #AI #AdaptiveAgent #BuildInPublic