Research

This page will hold original findings from real-world AI coding sessions: which practices correlate with stronger outcomes, where adoption stalls, and what actually changes in a team's first 90 days with agents.

Right now it holds none, because we haven't published any. We're pre-launch. What follows is what we're building, how we'll publish it, and the rules we'll hold ourselves to — written down before the first finding, so you can check our work against it later.

What this will be

Most writing about AI coding is opinion sampled from anecdote. The underlying sessions — the prompts, the plans, the rework, the abandoned attempts — are invisible in PR counts and seat-license dashboards. Accrete observes those sessions directly (that's the product), which gives us AI coding research data that mostly doesn't exist in public yet:

• Which concrete practices — planning before coding, delegating to subagents, task tracking, context management — correlate with stronger session outcomes, and which don't.
• Where adoption stalls: which practices teams pick up quickly, which plateau, and what separates people who use a tool from people who are genuinely proficient with it.
• How usage changes over a team's first 90 days — what the early weeks predict, if anything.
• Review and rework patterns in agent-written code: how often it ships as-is, how often it gets reworked, and by whom.

Findings, not opinions. If the data says something boring, we'll publish something boring.

Methodology and privacy commitments

Stated plainly, because this only works if you can trust it:

• Aggregate and anonymized only. We will never publish an individual's or a company's identifiable data. Each customer's data lives in an isolated per-company database to begin with; published research is a further aggregation step on top. (Full detail: privacy.)
• The method ships with the finding. Every published result includes how it was measured, the sample, and the known confounds. No naked percentages.
• We publish what the data honestly supports — including null results. "We expected X and found nothing" is a finding. Expect some of those.

Planned first drops

In progress, no dates promised:

1. Practice signals vs. session outcomes — which observable behaviors correlate with sessions that end well.
2. The adoption plateau — where teams stall after rollout, by practice.
3. The first 90 days — how a team's usage actually evolves after adopting agent tooling.
4. Rework in agent-written code — how much of it gets revised before merge, and what predicts that.

The public evidence so far

Until we have our own findings, here's the external research worth your time — with honest one-line summaries:

• METR RCT (July 2025) — experienced open-source devs were 19% slower with AI while estimating they were 20% faster. Scope matters: experienced devs, in mature repos they knew well.
• Google DORA 2024 report — AI adoption correlates with a 7.2% decrease in delivery stability.
• DX AI Measurement Framework — a sane structure for measuring AI assistants and agents instead of counting seats.
• DX Core 4 — the broader developer-productivity framework the AI measures slot into.
• GitHub Copilot longitudinal case study (2025) — a mixed-methods look at how Copilot use actually plays out over time, beyond week-one impressions.
• OpenAI's 2025 enterprise analysis — roughly a 4x productivity gap between power users and typical users with identical tool access.
• LeadDev Engineering Leadership Report (March 2025) — only 6% of 617 engineering leaders report significant AI productivity gains.

The pattern across these: tool access doesn't predict outcomes, self-reports don't match measurements, and the gap between best and typical users is the real story. That gap is what our data observes directly.

Get the drops

Research publishes to the early-access list first, before anywhere else. Join the early access list — or, if you want your team's (anonymized) sessions in the dataset and a say in what we study, become a design partner.