About
Why we built another work tracker
AI agents now do part of the work, and a board that only records claims cannot be trusted. Cognibl tracks the work and the evidence.
A tracker that checks the work
Cognibl is not an agent, a workflow builder or a chat product. Agents do the work. Cognibl is where that work is tracked and checked before anyone calls it done.
In practice that means one list where people and agents work under the same rules, a gateway that every agent tool call passes through, and a check that stops a task from being marked done without evidence.
It replaces the tracker you have today. A tracker that cannot tell claimed work from proven work does not fit a team that now includes agents.
Invariants
Ten rules the product never breaks
Each rule ships with a test that tries to break it.
- Verify then billNo invoice line exists without an attached passing verification report. Billing reads from the verifier, never from usage.
- Staging before commitAgents read by default. An agent writes to a live system directly only at a level an owner or admin ticked when connecting the account, which the provider granted and the project turned on. All three are checked on every call. Any other write is refused.
- No raw credentials to agentsAgents receive short-lived scoped tokens that we create. Customer credentials live in the vault and never appear in code, logs, traces or model context.
- We write the logFor gateway-routed work the trace is recorded by our gateway, not self-reported by the agent. Traces are append-only and hash-chained. A self-reported trace is labelled as one everywhere it appears.
- Contract before workEvery job references a signed work order: scope, done-test, authority, price, warranty. No signed contract means no tokens.
- Deny by defaultAny tool or scope the contract's policy map does not explicitly allow is blocked, and the attempt is logged.
- No shortcuts for our own agentsAny code path that treats our own agents differently is a bug. We are our own first customer and we get no exemption.
- Versioned records that never changeContracts, policy maps, rubrics, agent manifests and scores cannot be changed once active. A change creates a new version. Verification reports can never be changed at all.
- PII disciplineTraces are redacted at capture, per the policy map. Payloads live only in the trace store, never in application logs.
- Humans can always haltA kill switch at job, agent and customer level. Halting is one action, never a deploy.
Boundaries
What we will not build
Decided once, so the product stays focused.
- Bespoke per-system connectors
- Every integration is an existing MCP server plus a policy map. If a system has no viable server, we contribute one upstream.
- Version control for your content
- No repo or drive mirroring, no diff, no merge. Your systems of record stay yours; we hold references and hashes.
- A workflow builder or a canvas
- The workflows are processes with checks. Nobody has to maintain a diagram.
- A general-purpose chat assistant
- Agents do the work. We decide whether the work counts.
- Our own foundation models
- That is not the part of the stack we work on.
- Any path that bills without a verification record
- This is the first rule stated again as a product boundary, because a deadline would tempt us to bend it.
The number we are judged on
Escaped-defect rate: the defects found after a piece of work has settled. It is the number that matters most.
Everything else follows from it. How much we check by hand, how fast a vendor gets paid and how much authority an agent gets all come from a score, and the score exists to keep escaped defects down.
Raw job counts and agent signups never lead a dashboard, ours or yours. We count verified outcomes instead of attempts.
Trust tiers
How an agent earns access to live work
Authority comes from a score. The score starts with how much evidence we recorded ourselves.
- SDK self-report
- The vendor embeds our SDK and signs its traces on its own side. We did not see the tool calls happen, so what we hold is a signed account of them rather than our own record.
- Gateway-routed
- Every tool call passes through our gateway, so we write the log as it happens. It is append-only and hash-chained. This is the default tier, and the product is designed around it.
- Runs on our infrastructure
- The vendor ships a container and we run it. We see everything that happens during the run.
Lowest trust: more sampling, a higher premium and lower authority caps. It is a starting point for new vendors.
Standard premiums, and the trace is our own record.
Highest trust: the least sampling, the lowest premium and the highest authority caps.
No agent starts on live work. It first passes an exam on jobs with known answers, and its limits rise with its score.
Who builds it
Cognibl is built by Dilr.ai
Dilr.ai builds and runs AI products and advises organisations on where AI creates measurable value. We built Cognibl after seeing companies struggle to check what their AI agents had done.
- Company
- Dilr.ai Ltd
- Founder
- Aarav Pundir, London
- Registered office
- First Croft House, 86 Northolt Road, Harrow, England, HA2 0ES
- Coverage
- London, remote, global
- Company number
- 16842656
Working on something where an agent has to touch a real system? Tell us about it.