Skip to content

1.0Product

One tracker for people and AI agents

Every task has a written definition of done and the evidence behind it. Engineering leads and business owners see the same numbers. Agents work with a limited key, and nothing is marked done without proof. This page walks through the product one screen at a time.

Track people and agents in one place

Tasks, sprints, priorities and effort work the way your team expects, and AI agents are workers in the same tracker. When you sign in, you see your work grouped by status, with the day's refused status changes beside it.

1.1Tracker

Your work

4 open

Status is an icon, priority is a number, and you add a task by typing into a row. There is no pop-up form to fill in.

Priority runs from P1 to P5 and effort uses 1, 2, 3, 5 and 8. The burndown counts the same estimate you see.

Harness, Graph and Loop come with their statuses and moves set up, or you can draw your own. Each project picks one in settings.

No task is done without proof

Every task has a definition of done and a list of proof versions next to its scope. Cognibl checks the proof against the definition before the task can move to done. Only a passing proof moves it.

1.2Proof of work

Proof of work

INVO-13

definition of done · 4 checks

A correction creates a new version, and nothing is overwritten. The status moved because the proof passed.

The life of one task

Opened

An id and a definition of done, agreed before anyone starts.

Picked up

A person or an agent takes it. The agent's key opens only the tools the work needs.

Run

Every tool call passes the gateway. Refusals are traced too.

Proved

A CSV plus the coverage report, versioned on every correction.

Verified

Checked line by line. A failed line sends it back with the reason.

Done

The status moves and the task counts toward verified throughput.

A gateway controls what agents can reach

For an agent vendor, Cognibl is one MCP endpoint and a temporary key. Anything the key does not allow is refused. Every call and every refusal goes into an append-only trace, and the whole chain can be re-hashed to prove nothing changed.

1.3Gateway

Trace

chain intact4ad1efd015

all 135 requests re-hash in order, with no gaps

Payloads are redacted before they are written, and entries are never changed. Move the lens over the rows to read each reason.

Deny by default

Branch writesbranches and pull requests on linked repos
Library accessthis project's skills and agent manifests
Everything elsenot named by a key, not reachable

Each capability is a toggle that stays off until a person turns it on. What you see in settings is exactly what the gateway enforces.

Agents read your tools by default. An admin can let them write to Google Workspace directly, level by level. A staging queue that holds a write until it is approved is not built yet.

Metrics that count only verified work

Five main metrics, all based on verified work rather than activity. An AI layer reads the proof and writes the summary a stakeholder would otherwise ask a person for.

1.4Signals

Where time goes

median cycle time by stage

Splitting the cycle into stages shows when a long spec comes with a low first-pass rate. That pairing is the real problem.

First-pass rate

Agent tasks passing on the first try. A measure of spec clarity.

Human wait share

Time spent waiting on a person for review or a reply.

Verified throughput

Verified completions per week. It replaces velocity.

Reopen rate

Tasks reopened within 30 days of done, plus reopens from warranty claims.

AI reads the proof before you do

A summary of proof against the definition, on every status change.

A project summary that stays current

What moved and what is blocked, always current. Share a link.

Try Cognibl free

Free for one person. $5 per seat for teams.