Developers at All Day AI Agree: Coding Agents Need Strong Checks

A vendor-pitch-free conference with a shared playbook

According to Kilo Code, it helped run All Day AI on October 22. The free online conference had more than a hundred sessions and one hard rule: no vendor pitches. The recap, written by Federico Rubiano and dated November 2, 2026, draws on fourteen speakers from the developer track.

The speakers work at a stock exchange, a bank, a genomics company, a travel platform, a legal software company, a security vendor, a chip maker, Google, and several startups and consultancies. Despite those differences, Kilo says the same five practices came up across the talks. The part of the post available here details four of them.

Review is the bottleneck

Sachin Hiriyanna, a staff engineer at Navan, opened with public GitHub data: 43 million pull requests were merged in August 2025, up 23% on the year before, while comment activity on those PRs barely moved. His reading is that agents produce PRs faster than humans can read them.

His answer is to read what can hurt and let the agent read the rest. He keeps a checklist covering crashes, lost updates, silent fallbacks, widened permissions and untested branches, and sorts files by risk. Three sub-agents review the same PR, one against the checklist, one against the product requirements document and one against the design file. Their questions come back as one deduplicated list that he vets before it reaches the author. "The job is to approve on purpose," he said.

Shon Urbas of Pandium described the other side. His agent wrote tests that pass by construction, including assert True, and a human caught them. His team keeps two human reviewers per PR, and the second one still finds one or two real problems each time.

Put decisions in the repo and the checks out of reach

Parthiban Rajasekaran, a principal engineer at Capital One, described PRs that pass build, tests and lint yet break the system, because a design decision lived in a meeting and the agent only read the code. He named four drift patterns:

• Layer violations
• Parallel abstractions
• Bypassed standards
• Inconsistent failure handling

His fix is a one-page architectural intent file at the repo root, invariants encoded as checks that fail the build, and one added PR template question: which decision does this change touch? "The goal is not zero drift," he said. "It is drift caught while it is still cheap."

Max Artomenko at Illumina synced requirements and test cases from Jira, Jama and a quality management system into Markdown in the repository and pointed agents at them from AGENTS.md. On one mid-sized application, the agent's coverage check found seven gaps between requirements and tests and about 10 requirements the code only partly satisfied. He estimates the documentation time saved at roughly 80%, a figure Kilo labels as his own estimate.

Mradul Shrivastava of NASDAQ summarized his architecture as "the LLM drafts and everything else gets checked." Acceptance criteria are locked before code exists, and prompts, memory and model versions are pinned in Git. After three failed attempts the pipeline escalates to an engineer. A prompt version that scored 49 out of 50 on his golden dataset stayed blocked because the price calculation it touched requires 100%.

Viktoriia Voloshyna showed the failure mode in a live demo with Claude Code. The agent reported 93.7% accuracy on a sentiment classifier, but the real figure on the larger set was 0.684, because it had learned to tell drama from horror. Daniel Scheufler of Improving now writes tests with a separate agent, on a different model, from the one that writes the code.

The harness is where the leverage is

Ashish Kattamuri, a staff engineer at Proofpoint, explained why sessions degrade. The model starts from zero on every call, and the harness rebuilds the whole conversation each time. In his own session, about 120,000 tokens of a million-token window were used before his first message. He also said output tokens cost about five times input, and that recall drops for material in the middle of a long context.

Madeline Webster, a development manager at Clio, works mostly from the command line and uses plan mode first. Before a release she pointed agents at 40 to 50 minor bugs overnight, with draft PRs only, no merging, a maximum PR size and product questions marked blocked. The result was 12 to 15 PRs in about three hours for a couple hundred dollars, according to her account.

Chris Lucian of Hunter Industries runs agents for hours. He says unit tests, coverage, mutation testing, linting and static analysis form the feedback loop that lets an agent correct itself unattended.

What it means for teams using agents

The talks point to the same shift: the cost of shipping has moved from writing code to reviewing and verifying it. Teams in the recap responded with written intent in the repository, tests and gates that agents cannot edit, and a well-built harness around the model. Several of these measures, such as an intent file or a PR template question, require no new budget, as Rajasekaran noted.