Boden FullerAgentOps
Fresh-context validation for coding-agent work
Coding agents are stochastic and can declare work done when it is not. AgentOps turns one behavior into one bounded experiment, gives the exact result to a fresh validator, and stores the verdict under your control.
RPI -> Plan -> Implement -> fresh Validate -> durable verdict -> report and stop
Install
# Install the optional CLI.
brew tap boshu2/agentops https://github.com/boshu2/homebrew-agentops
brew install agentops
# Keep one canonical checkout and link it into every installed runtime.
git clone https://github.com/boshu2/agentops.git ~/.local/share/agentops
cd ~/.local/share/agentops
ao skills link
ao skills link creates source symlinks under the portable
~/.agents/skills root and every detected runtime skill root, including Claude,
Codex, Gemini/Antigravity, and Cursor. It never replaces a real directory or a
foreign symlink. Updates stay simple:
cd ~/.local/share/agentops
git pull --ff-only
ao skills link
Without Homebrew, build the optional CLI from the checkout and then link:
cd ~/.local/share/agentops/cli
go install ./cmd/ao
cd ..
"$(go env GOPATH)/bin/ao" skills link
The 3.x runtime plugin installers remain only as migration compatibility for this release. New installs do not need a plugin cache, hooks, or a runtime-owned copy of the AgentOps corpus. See the migration guide for removing an old plugin install.
Core workflow
> use plan for "rate-limit /login"
PlanPacket: normal + burst edge scenarios, exact scope, first acceptance check
> use implement
CandidatePacket: RED -> GREEN -> refactor, actual paths, content manifest
> use validate
verdict.v2: FAIL — burst refill violates scenario S2
checked: S1, S2, subject identity, write scope
not_checked: load behavior above declared limit
Or invoke rpi with rate-limit /login to run the three responsibilities once and
receive one report. RPI stops after PASS, FAIL, or NOT_PROVEN; the caller
decides whether to revise, deliver, or abandon the work.
Core skills
| Skill | Responsibility |
|---|---|
rpi |
invoke Plan, Implement, and fresh Validate at most once |
plan |
define one behavior, acceptance, evidence, and write scope |
implement |
run one bounded experiment and describe the candidate |
validate |
independently judge exact content and persist verdict.v2 |
learn is an optional later analysis of verdict collections. premortem,
postmortem, council, and idea genies are caller-selected strategies.
Factory/runtime skills such as NTM, Agent Mail, Gas City, and swarms are optional
adapters. None can change core sequencing or a verdict.
The evidence contract
A PASS binds:
- unchanged acceptance;
- a deterministic manifest of files, symlinks, deletions, executable bits, and content digests;
- complete changed-path coverage inside the Plan write scope;
- distinct author and validator context IDs;
- an explicit freshness attestation;
- criterion results, evidence references, checked scope, and omissions.
Missing identity, mutation, or incomplete coverage is NOT_PROVEN. A proven
out-of-scope change or failed acceptance criterion is FAIL.
Verdicts default to .agentops/verdicts/sha256/<digest>.json. They are plain,
content-addressed JSON and do not require Git, ao, a tracker, a hosted service,
or a provenance ledger.
Product boundary
AgentOps owns intent shaping, one bounded experiment, exact content identity, independent judgment, and the durable verdict. It does not own retries, budgets, queues, work claims, Git, CI, PRs, merges, closure, release, or delivery.
Use your repository's existing direct-push, PR, merge queue, hosted CI, and release process after validation. Local and cloud agents use the same packet and verdict contracts.
Honest status
Fresh independent judgment is a practical trust boundary, not a guarantee that stochastic output is correct. Context identities and freshness are declared facts, not cryptographic isolation. The longer-term learning hypothesis—that recurring verdict findings can improve future context and deterministic checks—remains off the critical path and must be measured.
Product boundary · Operating loop · CLI commands · Skill router · Docs index
Contributing: docs/CONTRIBUTING.md. License: Apache-2.0.