The SDLC defends against human failure. Your agents fail differently.
Models fail confidently, quickly, and at scale — a process built for forgetting and fatigue is not adequate for hallucination and reward hacking. ADLC gives every phase an explicit exit contract: deterministic gates leave machine-checkable evidence, and two human gates record attestation.
Install the toolkit and your agent integrations
curl -fsSL https://www.agenticlifecycle.ai/install.sh | shmacOS and Linux, Node 18+. Then adlc init in your repository. Cursor and OpenCode need one manual step, which the installer prints — see what gets installed. Windows is not supported.
In-session hooks are best-effort and harness-dependent. The commit-time CI gate is the unbypassable control.
Eight control points. A gate closes every one.
Each phase ends on an exit code: 0 passes, 1 is an operational error, and 2 means the gate refused. That is what makes the lifecycle executable in CI rather than aspirational in a wiki.
- P0Triageroutedadlc ticketPASS0
- P1Interrogatehuman gateA personATTEST
- P2Decomposecold-startadlc coldstartPASS0
- P3RailRED gateadlc rails-guardPASS0
- P4Buildgreen gateadlc build-gatePASS0
- P5Prosecutezero findingsadlc prosecutePASS0
- P6Integratehuman gateA personATTEST
- P7Distillfeeds the next runadlc lesson-foundryPASS0
Eight ways agents fail. None of them human.
Each one is answered by a specific gate at a specific phase, which is the only reason the register is worth writing down.
Attached to this record: the gate, the question, and the verdict.
Two of these are refusals. They are here because a record that only attaches its passes is not evidence, it is marketing — and because a gate that never refuses is not a gate. The commands and the exit codes are real; the verdict line under each is the record's own one-line summary of the result, not a verbatim transcript of what the tool prints.
$ adlc spec-lint ticket.md ✓ PASS: every acceptance criterion names a verification method
$ adlc rails-guard --base main --ticket T42 --tickets .adlc/tickets.json ✗ FAIL: a frozen rail changed
$ adlc hollow-test --test-cmd "npm test" --base main ✗ FAIL: one changed-code mutant survived
$ adlc review-calibration --review-cmd "adlc review --base {base}" --plants 3 ✓ PASS: planted-defect recall met the configured threshold
Native to the agent you already run.
Seven harnesses get a native integration. Everything else installs the skills catalog, which is skills only — no hooks, MCP, agents, or rail enforcement.
Claude Code
Phase-routing skill, slash-command gates, rails-guard and build-gate hooks, and a multi-lens prosecutor panel. Claude is the model via --prompt-only.
Codex
A native Codex plugin with phase routing, frozen-rail feedback, gate tools, and project review agents.
Cursor
Marketplace plugin with sessionStart context, preToolUse rails, packaged prosecutor agents + MCP wrapper (channel unverified), /adlc-* commands, and CI as the backstop.
OpenCode
Enforce-by-default rails-guard, full /adlc-* lifecycle commands, native adlc_gate/adlc_prosecute tools, and a multi-lens prosecution loop with a verifier.
Pi
Proactive tool_call blocking and reactive snapshot revert, five phase skills, native adlc_gate/adlc_prosecute tools, and the strongest team-install story among the harnesses.
Google Antigravity
Advisory PreToolUse rails-guard for agy, plus skills and a prosecutor agent. The unbypassable CI rail-freeze gate is the real control.
GitHub Copilot
A native Copilot CLI plugin: seven lifecycle hooks, six phase skills, allowlisted MCP gate tools, and six read-only prosecution agents.
Spend heavy at the ends, light in the middle.
The unit of account is cost per merged, verified change. Not tokens per developer per month.
This is the SDLC inverted, because misbuilding is expensive and building is cheap. Review stops being the bottleneck, and it goes deeper than a human team could sustain anyway. Every merge carries an evidence manifest, so auditors get an answer without an engineer translating. Seniortime goes to the two human gates instead of rubber-stamping diffs. And each run costs a little less than the last, as findings distill into lints and skills nobody has to learn twice.
Rolling this out across an org?
Unreviewable agent output is an audit problem, not just an engineering problem. ADLC produces a gate-by-gate evidence trail your auditors can actually read, and 6 of the 8 gates need no human in the loop to produce it.
Is this what I meant?
Is this what I meant, running?