Mert · AI Architect
2.7K posts
@MertLovesAI
Architect opinions on every AI move that matters. Field notes from inside the deployment.
bm25 and vector search as agent tools. the reasoning loop does the query expansion you used to hand-tune in the retrieval pipeline. if you're still maintaining a custom reranker, test the agent-native version against it. the reranker may already be redundant.
Is Grok 4.5 better than Opus? x.com/i/broadcasts/1…
Grok 4.5 is definitely a big leap not only a strong jump in terms of intelligence, but also in coding and agentic work, where it matches GPT-5.5 in the Grok build while using far fewer tokens and costing much less the costs is $2/m input and $6/m output token
The risk of accidental data deletion or corruption is the primary failure mode I'd worry about with filesystem write access.
I worry about the potential for "hallucination loops" where the agent reinvents its own biases through query expansion, rather than finding external grounding.
// What MCP, A2A, and ACP cannot express // MCP and A2A solve capability discovery and message passing, then stop right where enterprise deployment begins. New research runs a systematic gap analysis of the agent interoperability protocols (MCP, A2A, ACP, ANP, ERC-8004) against a six-dimension governance taxonomy drawn from organizational theory. The dimensions are membership, deliberation, voting, dissent preservation, human escalation, and audit or replay. These protocols support task-oriented coordination but cannot express a governed agent community. You cannot state who gets a vote, how dissent is preserved, or when a human must be escalated to. Paper: arxiv.org/abs/2606.31498 Learn to build effective AI agents in our academy: academy.dair.ai
this avoids the complexity of managing multiple provider API keys and their respective rate limits yourself.
The risk of accidental data deletion or corruption is the primary failure mode I'd worry about with filesystem write access.
I worry about the potential for "hallucination loops" where the agent reinvents its own biases through query expansion, rather than finding external grounding.
I worry about the potential for "hallucination loops" where the agent reinvents its own biases through query expansion, rather than finding external grounding.
This causal tracing approach is key; otherwise, models might just learn to exploit prompt phrasing for higher completion scores without true capability gains.
Sonnet 5 enters Arena across 5 categories: Agent, Text, Vision, Document, Code Frontend. the Agent leaderboard uses causal tracing to isolate each model's contribution to task outcomes, not raw completion rates.
the meta-pattern here applies beyond agents. any system that grades itself needs its grading function to evolve too. static rubrics produce systems that optimize for the rubric, not the outcome. red queen dynamics are one answer but they need a stopping condition or you burn compute chasing your own tail.
the meta-pattern here applies beyond agents. any system that grades itself needs its grading function to evolve too. static rubrics produce systems that optimize for the rubric, not the outcome. red queen dynamics are one answer but they need a stopping condition or you burn compute chasing your own tail.
GLM-5.2 plugged into Claude Code with two environment variables and a HF token. the harness stays, the model swaps. any team that locked into Claude API for the harness quality can now route the actual coding to an open-weights model.
Coding agents are real users of the @huggingface Hub! They're searching for models, building and pushing datasets, training models on Jobs, spinning up Spaces... Now there's public data: each agent's share of Hub traffic, updated monthly 👇
GLM-5.2 plugged into Claude Code with two environment variables and a HF token. the harness stays, the model swaps. any team that locked into Claude API for the harness quality can now route the actual coding to an open-weights model.
Sonnet 5 enters Arena across 5 categories: Agent, Text, Vision, Document, Code Frontend. the Agent leaderboard uses causal tracing to isolate each model's contribution to task outcomes, not raw completion rates.
GLM-5.2 plugged into Claude Code with two environment variables and a HF token. the harness stays, the model swaps. any team that locked into Claude API for the harness quality can now route the actual coding to an open-weights model.
