Autonomous coding agents hit production: 87% success rate, 10x faster delivery
A Tencent production system called SiriusDeliver is now running 18,240 autonomous data warehouse sessions per month with an 87.2% success rate and no human intervention, according to research published in September 2026. Median delivery time dropped from 228 to 23 minutes, and engineer effort fell from 95 to 11 minutes per task. The shift from AI assistants to autonomous agents is reshaping how software is built, tested, and deployed.
Bottom line — Tencent's SiriusDeliver cuts engineer effort by 88% per task, with 87.2% autonomous success rate.
Go deeper (8)
- The agentic coding era is defined by agents that can plan, write, test, and debug code across multiple files without human prompts, per multiple research papers from Google, MIT, and Stanford published in September 2026.
- A new framework called Spec-Driven Agentic Development (SDAD) treats specifications as contracts, with measurable acceptance criteria, to reduce what researchers call the 'Ambiguity Tax' — the hidden cost of imprecise requirements, per the DEV Community article citing Bhati's research.
- The Verification Tax is the hidden cost of reviewing agent-generated code: teams must invest in automated verification gates to avoid drowning in rework, according to Jarmak's review of 264 works on reliable coding agents.
- At Nutrient, a software company, agentic AI usage is now mandated — engineers write code manually only as a fallback, per a CIO opinion piece. The company reports that impact-weighted pull requests nearly doubled in a few months.
- Dedicated testing agents are emerging as a separate concern: 'The author cannot be the examiner,' argues devassure.io, warning that agents testing their own output create a blind spot similar to developers testing their own code.
- The τ^τ-bench benchmark shows that the best coding agents pass only 23.9% of 'build an agent' tasks against a human expert ceiling of 82.2%, per the DEV Community article. Agents optimize for visible tests, not for behavioral correctness.
- Open-source agent frameworks like OpenHands and CrewAI now allow teams to build custom multi-agent workflows with specialized roles (planner, coder, reviewer), per thelinebuild.com's practical guide.
- Qwen3-Coder-Next, an 80B-parameter model released in early 2026, achieves near-top-tier performance while running on consumer hardware, per Talking Tech, making powerful local agentic coding possible.