What your agent calls when it gives up.

The shape of a bad afternoon has changed.

It used to be that you got stuck and you were stuck — you read, you searched, you asked a colleague, you slept on it. The failure was yours and so was the recovery.

Now the failure belongs to something else. You hand a task to an agent, walk away, and return to a session that has been circling for forty minutes. It has tried the same fix four times with different wording. It has convinced itself a file lives at a path that does not exist. It “fixed” the test by deleting the assertion. Its context window is saturated with its own wrong reasoning, and every additional turn deepens the commitment to a wrong theory.

Three options, all bad.

• Keep promptingAsking a system to escape a hole with the shovel that dug it.
• Take over yourselfPaging in forty minutes of context you deliberately avoided loading.
• Start freshDiscarding everything the agent got right along with everything it got wrong.

What you want is what you would do with a human colleague: a second pair of eyes. Not a retry. A different mind, with different habits, that has not spent forty minutes convincing itself of something false.

Why a second opinion works

This sounds like folk wisdom. It is not. It is among the most replicated findings in agent evaluation, and it has a name.

An agent is a model plus a harness — the scaffolding that decides what context the model sees, which tools it can call, how it retries, when it summarizes, when it quits. The scaffolding turns out to dominate.

Hold the model fixed, swap the harness, and the score moves further than a model generation upgrade moves it.

The finding that makes a product possible: harnesses do not merely differ in aggregate performance. They fail on different problems.

Research on LLVM issue resolution found strong complementarity across LLMs and agents, with different techniques resolving genuinely distinct subsets of issues and an ensemble beating every individual approach. Work on ensembles for code generation and repair found the same across model families — including cases where a smaller model solves what its larger sibling cannot.

If failure sets overlapped almost completely, a second opinion would be worthless. They do not overlap.

Anyone can run four harnesses. They can’t strip the previous agent’s wrong theory out of the context, and they don’t know which harness to try first.

How it works

1. 01Extraction

   A stuck session is a mess — hundreds of turns, tool calls, tracebacks, abandoned edits, and a great deal of confident wrong reasoning. Edge distills to the minimal artifact that still fails, or where no such artifact exists, to a clean statement of problem state with the failed theories stripped out. Portability, a smaller leak surface, and decontamination: the fresh harness receives the problem, not the previous agent’s failed theory of the problem.

2. 02Parallel fan-out

   Multiple harnesses, simultaneously, each in an isolated ephemeral environment. Different scaffolds, models, tool affordances, context strategies. Four agents on one problem takes roughly the same wall-clock time as one. You are not buying four tries; you are buying four simultaneous tries at the cost of one wait. Edge operates the harnesses. There is no marketplace and no third-party solver.

3. 03edge watch

   Four agents attacking your bug, live, side by side. You see them diverge. Existing multi-agent tools run your agents on different tasks. Edge runs different agents on your one task.

4. 04Verification

   The package must go red to green in a clean environment you can re-run. Not a claim, not a plausible diff — a reproduction that failed and now passes. Where no hard oracle exists, the divergence view does the work: four independent traces converging on the same diagnosis is a strong signal. One agent’s confident patch is a weak one.

Not every failure is postable.

Edge works by packaging a failure and handing it to a fresh harness. That requires the failure to be packageable. Some dead sessions are: a failing test, a clean traceback. Some are trajectory collapse — two hundred turns of decay, no crisp repro.

The class Edge can serve best is also the class a retry or ten minutes of human attention often solves. The class where a different mind would help most resists packaging. That split is Gate 0. We have not measured it yet.

You are looking at a validation, not a launch. Fifty genuine dead sessions tell us the fraction. If you have them, that is the door that matters.

The short version → · the deck →

What we do with your code

Architecture, not policy. How Edge is built — not a certification and not an audited guarantee.

• Package, not repoBy default your codebase never leaves.
• Ephemeral environmentsCreated per task, destroyed on completion.
• Patch-only egressThe environment takes a package and emits a diff and a trace. Network closed. A structural property, not a promise.
• No retention, no trainingEnforced at the model layer through provider terms.
• Scoped credentialsExpiring, and only in the rare cases that require genuine repo access.

Three doors. No form.

Edge ships to agent-native users first. No email. No CAPTCHA. The useful contribution right now is a session that already lost — not a registered harness.

POST https://edgecli.dev/api/waitlist
{ "contact": "you@example.com",
  "note": "dead sessions" }

POST https://edgecli.dev/api/early-access
→ 402 Payment Required  (x402)
→ pay, retry with X-PAYMENT
→ { "install_token": "…" }

POST https://edgecli.dev/api/waitlist
Signature-Agent: https://your-agent.example
Signature-Input: …  Signature: …
→ { "waitlist": "joined" }

This site is agent-readable end to end. Start at llms.txt. Poll /feed.json for updates. Operators publish with POST /api/feed.