Skip to content

CONCEPT Cited by 2 sources

Agent as first-pass investigator

Definition

Agent as first-pass investigator is the operational model in which an AI agent is delegated investigation, diagnosis, and a proposed fix as its bounded scope of authority — the agent prepares a draft PR / report / proposed change, but a human reviews and merges. The agent is not authorised to merge, release, or otherwise commit the change to production state.

Atlassian's 2026-06-01 Jira-team post gives the canonical wiki articulation:

"The key is that the agent handles the repetitive first pass: investigation, diagnosis, and a proposed fix. Engineers validate the change before it is merged. What used to require hours of manual investigation can now become minutes of review." (Source: sources/2026-06-01-atlassian-how-we-cut-up-to-80-of-engineering-chores-using-ai-agents-in)

The three-stage scope of authority

The agent's bounded scope, as defined by Atlassian's KTLO workflow:

  1. Investigation — read the work item, fetch logs / call stacks / failing-test outputs, query the codebase. Pure-read operations. No state mutation.
  2. Diagnosis — classify the issue (real vs false positive, unit vs integration vs visual flake, repo with skill vs without, …). Decision recorded as a comment on the work item.
  3. Proposed fix — apply the relevant skill's fix pattern, commit changes to a branch, open a draft PR. PR contains the diff and an explanation; PR is draft so it cannot be merged without human action.

Out of scope: merge, deploy, close ticket, configure-flag-state in production, modify CI / repo settings.

Why human-in-the-loop is the load-bearing constraint

The cost economics of KTLO automation depend on the first-pass investigation cost being collapsed from hours to minutes while the review remains human. Atlassian's framing is explicit:

"What used to require hours of manual investigation can now become minutes of review."

If the human were removed (agent merges its own PRs), the saving would be larger but:

  • Review-fatigue protection vanishes — the agent's failure modes (false-positive triage, stale skill applied to wrong repo, regression introduced by fix-pattern misuse) ship to production silently.
  • Merge gate becomes the only safety net. Merge queue CI catches some classes of regression but not all (semantic bugs, unintended behaviour changes, partial refactors).
  • Authorship ambiguity. A revert / post-mortem becomes harder when no human was on the merge.

The "first-pass investigator" framing is the explicit choice to trade some agent throughput for review accountability.

Agent's exit conditions per stage

Atlassian's flaky-test workflow exemplifies bounded exits:

Stage Exit options
Triage (a) False positive → comment on work item, STOP. No PR. (b) Reproducible → enter fix stage.
Fix (a) Fix produced → open draft PR, comment on work item linking PR. (b) Could not reproduce / could not fix → comment on work item with diagnosis, STOP. No PR.

"If it looks like a false positive, the agent can stop and summarise that outcome, commenting on the original Jira work item. That way the engineer reviewing the ticket can quickly get a sense of what the agent did and what the agent discovered without digging." (Source: sources/2026-06-01-atlassian-how-we-cut-up-to-80-of-engineering-chores-using-ai-agents-in)

The bounded-comment exit is doing two things: (a) preventing "agent opens a PR for a false positive" failure; (b) returning diagnostic value to the human even when the agent can't act.

Sibling pre-human-review patterns

This concept composes with several wiki primitives that thicken the pre-human review step:

  • patterns/pre-human-agent-review (Atlassian Fireworks, 2026-04-24) — three-tier stack: adversarial sub-agent → CI / merge-queue gate → human architect. This concept names the first-pass / merge-gate split; pre-human-agent-review fills in the "adversarial sub-agent" stage between agent and human.
  • patterns/ci-as-agent-quality-gate — CI is the automated half of the gate. Agent reads CI output and patches before requesting review.
  • patterns/agentic-pr-triage (Fly.io / Phoenix.new) — the greenfield-issue-triage variant; same agent-drafts-PR shape but applied to feature work rather than KTLO.

Distinction from fully-autonomous agent

The "first-pass investigator" posture is deliberately less ambitious than fully-autonomous coding agents:

Dimension First-pass investigator (this concept) Fully-autonomous agent
Scope Investigate + diagnose + draft fix Investigate + diagnose + fix + merge + deploy
Merge gate Human review of draft PR None / agent-self-approves
Failure containment Human catches at PR review Merge queue + canary + revert
Throughput Review-bound Compute-bound
Best fit KTLO categories with high pre-existing pattern recognition Greenfield work with novel design risk
Risk profile Conservative; mistakes caught before production Aggressive; mistakes caught after production

For KTLO categories where the team already has years of pattern recognition, the first-pass-investigator model is the right posture — review is cheap (the team knows what good looks like), so the human-gate cost is low and the merge accountability is preserved.

Why "minutes of review" is the load-bearing economic claim

The 80% time-reduction claim only works if review of a draft PR really is minutes, not hours. That requires:

  • The agent's fix follows a known good pattern the reviewer recognises.
  • The agent's commit message / PR description explains the diagnosis clearly.
  • The agent doesn't open false-positive PRs — review fatigue is the failure mode that destroys the economics.

The triage-vs-fix split, the bounded false-positive exit, and the specialist-skill dispatch are all in service of keeping review-time low.

Seen in

Last updated · 542 distilled / 1,571 read