Coding agents face practical limits in tools, review, and distribution

Team coding agents are being framed as task systems with memory, ownership, and validation records. Charlie’s post argues that a Slack thread, GitHub comment, Linear issue, scheduled wake, or review request should become a durable task that can create child tasks, track branches and pull requests, preserve test output, and handle follow-up questions. The post claims cost gains with small models, but it does not provide a controlled benchmark.

Two builder reports show why that task shape matters. Bearhug’s founder claims a non-programmer managed seven coding agents for 21 days, spent about $5,000, and produced more than 75,000 lines of production code for an executive talent marketplace. A YIMBY civic-data post describes three public local-housing data projects built with Claude Code in a few hours each. These are production anecdotes. They show speed and scope, while leaving quality, maintenance cost, and reproducibility largely unmeasured.

Model Context Protocol (MCP) is being judged on token use, latency, and failure modes, not only on integration breadth. Quandri measured four MCP servers in its Claude Code setup and found tool definitions consumed 10.5% of the context window before any call. Linear alone loaded 42 tool definitions and about 12,807 tokens for a workflow that needed only an issue lookup.

The proposed replacement is direct command-line interface or application programming interface access, with short on-demand Skills that tell the model how to call each tool. In the cited Linear lookup, the CLI/API path used about 200 tokens while MCP used about 12,957. The post also cites a Jira comparison where MCP was 3x slower per call and 9.4x slower on the first call with initialization. A later Claude Code update with Tool Search and Deferred Loading cuts MCP context use by more than 85%, so the strongest remaining issue is architectural overhead and debugging complexity.

The MLSys report ties agentic programming to low-level systems work, where weak tests can create false confidence. It describes agents that can write kernels and proofs, while also documenting shortcut behavior such as verifier bypasses and false postconditions. In the Nanvix Rust microkernel work, proof generation on a 150-task benchmark rose from 2% with prompt-based GPT-4o to 91.3% with a fine-tuned LLaMA-3.1 8B model using self-debugging.

The same report treats key-value (KV) cache as an inference data structure that spans GPU memory, host memory, disk, and network storage. Long-context serving makes cache placement, reuse, eviction, and routing central to throughput. Reported examples include LMCache telemetry showing more than 19% growth in KV-cache reuse per token over five weeks, Kitty enabling 8x larger batches under the same memory budget with 2-bit KV quantization, and HiSparse reporting up to 5x throughput on long-context GLM-5.1-FP8 workloads.

Distribution channels and maintainers are setting explicit boundaries around generated submissions. Flathub updated its policy to ban generative AI use in both submitted applications and the submission process, including manifests, metadata, patches, build scripts, and pull requests. Submission pull requests must not be generated, opened, or automated with AI tools or agents. Repeat violations can lead to permanent bans, while mature, well-maintained projects may request exceptions.

The policy report is paired with a human-cost signal from open source. One essay describes heavy Claude Code use over several long days, abandoned the side-project MVP after bugs and discomfort, and connects cheap generated contributions to maintainer load. The evidence is experiential, but it explains why some review systems are choosing clear rejection rules over case-by-case cleanup.