$ ./agent-harness --pitch

THE AGENTHARNESS

A skills framework + a risk/trust harness for the Claude Code CLI.
Brainstorm → ship — with the right amount of care at every step.

press or Space to advance · F for fullscreen

01 / the problem

AI coding agents fail in two opposite ways

Vibe coding

Reckless autonomy — the agent executes fast, skips validation and understanding, and you inherit a diff you can't trust.

🐌

Analysis paralysis

Micromanaged into uselessness — asked to confirm every step, the speed advantage of an agent evaporates.

Both come from the same gap: the agent enters “with only a chat prompt and a shallow snapshot of files.”

02 / the real bottleneck

“Coding agents do not only need better prompts.
They need better repositories.”

The repo itself lacks agent-addressable context — so constraints get lost to chat history, validation is skipped, and architecture decisions get re-litigated every session.

03 / the reframe

“The app is what users touch.
The harness is what agents touch.”

Build a thin, git-operable control layer for the agent — classification, proof, and guardrails it can read and act on.

04 / the one principle

Two independent dials

RISK
↓ decides
how much proof & process a change carries
AMBIGUITY
↓ decides
whether a human is asked

Ceremony scales with risk. Interruption scales with ambiguity — not risk.

05 / two layers

The engine and the harness

The engine — skills

Invocable /skills that do the work: brainstorm · research · plan · build · review · ship. Each with hard gates and a defined handoff.

🛡

The harness — risk/trust

Decides before the engine runs: how much process a change needs, and when to ask a human. Intake · hooks · rules · evidence.

Engine answers “how do I build this?” · Harness answers “how careful, and who approves?”

06 / the workflow

brainstorm ship

/feature-intakeclassify risk lane + confidence · runs first
/brainstorming → /xia2explore intent · research what already exists
/writing-plansturn design into parallel-safe XML tasks
/using-git-worktreesisolated branch + worktree
/subagent-driven-developmentbuild task-by-task · two-stage review · delegates both whole-diff passes
/correctness-reviewadversarial bug hunt over the whole diff · findings scored (threshold 80) · standalone too
/intent-reviewthe last gate — diff vs the original request, blind to the plan
/compoundcapture non-obvious learnings
/finishing-a-development-branchPR · review · merge 🚀

Intake decides how many of these steps actually run.

07 / the router

“The human does not need to classify risk.
The harness does.”

/feature-intake runs a 10-flag risk checklist + hard gates → a lane, and a confidence score → whether to ask a human. Deterministic, every time.

08 / three lanes

Ceremony, sized to risk

tiny
direct edit · no plan · no human gate · hooks are the net
normal
plan + two-stage review · notify-and-proceed
high-risk
full chain · recorded rollback · human gate on ambiguity / hard gate

Proof is required, not asserted — and it scales down for small work.

09 / hard gates

These always force high-risk

auth
authorization
data-loss / migration
audit / security
external provider
public contract
weakening validation
high-blast file

Only a human narrowing scope can lower them — and a gate found mid-task escalates regardless of the original lane.

10 / enforcement over convention

Rules enforced by code, not hope.

What can be mechanized is a hook — fail-closed. A commit-time check re-reads the diff and corroborates the declared lane: the agent can't label a risky change “tiny” and slip it through.

secrets scan · debug-artifact check · targeted tests · risk corroboration · blast-radius · ruff

11 / why you can trust it

Trust comes from evidence, not assertion

L1
Evidence over assertion — claims backed by runnable artifacts (test exit code, output), not self-report.
L2
Enforcement over convention — gates are hooks, fail-closed.
L3
Reversibility — worktrees, per-task commits as rollback points; irreversible ops STOP.
L4
Verifiable classification — self-declared lane cross-checked by a hook and an independent reviewer.
L5
Independent review — spec + quality reviewers see diff & plan only, never author context.
L6
Confidence + escalation — scored confidence; escalations logged, default-deny on no response.
12 / it compounds

Every session makes the next one easier

PLAN
WORK
REVIEW
COMPOUND

A knowledge graph gives structural context; /compound captures every bug, decision & insight into docs/solutions/ for future agents.

13 / what you get

Humans steer. Agents execute.

Less ceremony on small work

Typo / copy / narrow edits skip planning entirely.

More proof on risky work

Auth, migrations, contracts → full plan + reviews + rollback.

Asked at the right moments

Approval gates on ambiguity, not on every step.

Claims backed by evidence

“Done” needs a re-runnable verify artifact.

The human moves from per-task gatekeeper to rule author & exception handler.

14 / see it live

Try it yourself

An interactive guide — with a live /feature-intake classifier you can play with.

install: curl … install-harness.sh | bash -s -- --yes

15 / standing on the shoulders of

superpowers

@obra — skill engine

GSD

@gsd-build — spec-driven dev

Compound eng.

Every — learning loop

harness-experimental

@hoangnb24 — risk/trust

Thanks. Now go ship — gated by risk.

github.com/minhtran3124/agent-harness

⌂ home
▸ arrows / space to move · F fullscreen · swipe on mobile
1 / 16