AI News Briefing — Harvard study finds coding agents add code, not output

AI-assisted SDLC

[2026-10-09] Harvard researchers Fiona Chen and James Stratton used Jellyfish data on more than 700 software firms, from 2021 through March 2026, to measure what coding agents add. Adoption raised lines of code by 30%, commits by 20% and pull requests by 23%, but the rate of resolved Jira issues and epics did not change significantly. The working paper finds the gain absorbed by review: time from pull request to merge grew 49%, the share of PRs needing changes nearly doubled, and AI wrote only 23.3% of review comments. (source)

The output stalls at review. A team adding agents without adding review capacity should expect a longer queue rather than more shipped features.

[2026-10-09] Microsoft launched Microsoft-Decision-1 in Foundry, a decision model post-trained on Alibaba’s Qwen3.5-9B that scores a fixed set of options and returns calibrated probabilities as JSON instead of text. It costs $0.042 per million input tokens with free output, the same as TypeSafe’s Jev, accepts up to 32,768 tokens of text only, and has no open weights. Its speed claims against GPT-6 Sol come from Microsoft’s own internal teams. (source)

Microsoft’s own Foundry documentation tells customers to validate calibration on their own data, which is the test to run before letting it make an allow-or-block call.

For ML / Data Engineers: A calibration claim is only as good as the labels it is checked against. Hold out a few hundred past decisions with known outcomes, bucket Decision-1’s probabilities, and confirm the 0.8 bucket is right about 80% of the time before picking a cutoff.

[2026-10-09] Decision models are settling on one interface: TypeSafe’s System One schema, with choice, score and yes/no question types in a single request. Upstage’s Solar Decide, AWS’s open-weight Strands Decider and Ollama 0.35 all answer at /v1/systemone, and OpenRouter offers a TypeSafe-compatible route beside its own Decisions API. OpenAI has not signed on; its Decisions API, built on GPT-6 Luna, keeps its own shape. (source)

Code written against the shared schema can switch providers by changing a base URL. A portable request does not make two models’ probabilities mean the same thing, so thresholds still need re-tuning per model.

For Solution Architects: A team picking a decision-model provider today can write against /v1/systemone once and keep Solar Decide, Strands Decider and a local Ollama 0.35 as interchangeable backends. OpenAI’s Decisions API is the one that needs its own adapter.

[2026-10-10] Copilot for JetBrains adds a setting that stops MCP servers starting automatically for Copilot and Claude, lets enterprise admins set the default agent model, and drops support for IDE 2025.1. (official)

Anyone still on a 2025.1 IDE is now outside the supported range.

For Security Engineers: Switch off automatic start-up for MCP servers in Copilot for JetBrains, so no configured server runs for Copilot or Claude until a developer starts it on purpose.

[2026-10-09] Postman’s write-up of Agent Mode on Amazon Bedrock puts a number on tool overload: selection errors rose once an agent saw more than about 40 tools. Its fix is a root agent that searches tool embeddings to cut 170-plus tools to about 15 per request, then hands them to a context-isolated sub-agent. Prompt caching uses a 1-hour checkpoint for the fixed core, with a 5-minute one for changing context placed after it. (official)

Postman measured the 40-tool threshold on its own setup, but counting the tools your agent sees on each request takes a minute.

[2026-10-09] Ai2 replaced priority-based GPU scheduling, under which nearly every job ended up marked high priority, with GPU-time budgets that managers split across programs and projects, fair-share over a 7-day window, and an 8-hour protected minimum runtime. Over 30 days teams got 98% of the GPU hours they were owed, occupancy held at 98%, and p90 queue wait for debug jobs fell from 2 hours to 30 seconds. (official)

Budgets turned preemption into a rule instead of an on-call judgement. Interactive sessions lost out under the 8-hour cap, which Ai2 plans to fix with a separate CPU-only dev cluster.

Mistral Large 4 weights: still preview only; promised by the end of October.
Step 5 Preview’s weights: not on Hugging Face yet; due October 15.
Decisions API general availability: OpenAI’s API remains in public beta with no date.