The existential panic cycle around artificial intelligence has hit peak fever pitch again. Between frontier lab manifestos calling for mandatory capability pacing and papers asking why agents lie, cheat and coordinate in sandbox simulations, you would think AGI was about to break civilisation by lunchtime next Tuesday. Meanwhile, back in the actual engineering trenches where people build software for a living, developers are wrestling with far more earthly crises: coding agents running 25 times more continuous integration jobs, context windows collapsing under bloated markdown files, and agents burning forty dollars in API calls trying to parse a CSV because someone forgot a trailing comma.
There is a widening gulf between boardroom existential dread and daily production reality. Frontier labs might be worrying about runaway superintelligence, but anyone deploying agents to real users knows that reliability is still won in the boring details: schema validation, test impact analysis, deterministic error recovery, and sensible context isolation.
In this edition of AI++ we explore why agentic workflows are breaking traditional CI pipelines, get hands-on with managed harnesses and MCP elicitation, and find out what happened when an AI model decided to hack a piano.
Erick Ramirez | Apache Cassandra committer & Developer Advocate at IBM/DataStax
🛠️ Building with AI, Agents & MCP
Agentic coding and CI bottlenecks
If your engineering team has started giving coding agents free rein over pull requests, you have probably already noticed the collateral damage: your continuous integration pipelines are melting. Will Depue and the engineering team at Anthropic shared how internal agentic coding caused a staggering 25x increase in CI workloads over six months, completely crushing their monolithic test runners. Upgrading machine sizes only postponed the inevitable; the only durable fix was adopting graph-based Test Impact Analysis to run only the tests affected by an agent's specific diff.
The problem is that autonomous agents do not think like human developers when writing software. As Rachel Laycock notes on Martin Fowler's blog, agents do not suffer from fatigue, but they do generate structural bloat, duplicate abstractions, and subtle regressions that slip past naive unit tests. Earendil's SlopCodeBench benchmark quantifies this exact phenomenon, measuring how AI-generated code accumulates semantic erosion and ghost dependencies over time. If we are going to let agents write the majority of our code, our verification and CI infrastructure must become significantly smarter than the models generating the pull requests.
Harnesses over raw loops
The consensus around building production agents has decisively shifted from clever system prompts to robust, defensive software harnesses. The team at monday.com shared a practical case study on their feedAgent harness, illustrating why model intelligence alone is never sufficient. In high-volume production, reliability comes from strict input sanitisation, schema validators, hallucination tripwires, and hard execution bounds that prevent recursive runaway costs.
Context architecture is the other half of the battle. Harrison Chase and the LangChain team recently broke down how multi-agent systems should manage context, arguing that blindly forking a supervisor's entire conversational state into every subagent is a recipe for context compaction failures and wasted tokens. Instead, isolating subagent context and selectively injecting procedural memory (such as skillmem for MCP) keeps execution fast and deterministic.
Managed protocols and tool elicitation
Connecting agents to tools and external identities is finally maturing into standard platform infrastructure. LangChain announced native MCP support via langchain.mcp built on FastMCP for the 2026-07-28 specification, introducing interactive user-elicitation as native LangGraph interrupts. This means when an MCP tool needs human confirmation or missing credentials, the agent loop pauses cleanly instead of failing mid-turn.
At the same time, LangChain’s new Connections feature tackles the thorny problem of per-caller identity in multi-tenant agents, allowing subagents to act strictly under individual user OAuth credentials rather than a shared service key. Combined with OpenAI’s managed Agents API, cloud providers are turning agent orchestration, sandboxing, and session persistence into standard backend primitives.
🧠 New models
- DeepSeek v4.1-Flash is a 763B parameter causal encoder–decoder model with vision that dramatically shrinks KV-cache memory requirements, making long multi-turn agent sessions noticeably faster and cheaper to run.
- GPT-6 Astra is OpenAI’s latest frontier release tailored for computer use and multi-step reasoning. It carries a higher per-token price, but requires fewer aggregate planning steps to finish complex workflows.
- Meta Muse and Muse Spark 1.3 provide open-weight autonomous agents and frontier-level distillation at a steep discount compared to proprietary alternatives.
- TabPFN-3.5 is Prior Labs’ tabular foundation model that beats gradient boosted trees across standard benchmarks without requiring manual feature engineering.
- GPT-Live-1 brings low-latency, full-duplex voice interaction directly to OpenAI’s API with built-in telephony protocols and custom voice matching.
🗞️ Other news
🧑💻 Code & Libraries
- Amber Agent is a lightweight, open-source terminal coding agent built with only four dependencies and zero UI framework overhead
- Senro is an automated testing and reliability suite for validating WebMCP implementations across real browser environments
- doc-audit is a CLI verification tool that checks whether claims in
CLAUDE.md and AGENTS.md match actual repository code
- dbt Charts is a declarative YAML specification for building embeddable, SQL-backed charts inside AI chat interfaces
- Cline Desktop is an open-source workspace for running parallel open-weight models with MCP server integration
🔦 Langflow Spotlight
Langflow has opened applications for its new Langflow Champions program, inviting 25 community builders into early access for the upcoming major architectural release. The next version brings deeper multi-agent canvas orchestrations, native MCP server discovery, and enhanced state persistence components. If you are building production agent workflows visually and want direct access to the core engineering team to shape what comes next, apply to join the early access cohort.
🗓️ Events
David Gillardi and Tejas Kumar are back livestreaming the next build session of Walfly, an open-source wearable AI app built with IBM Bob. Now that the app is stable across both web and mobile after wrestling with Expo, this episode dives into handling audio recording checkpoints during long sessions and introducing new tools for managing background agents. Tune into the livestream on YouTube to see how the build comes together live.