Make AI coding agents observable, correctable, and cheaper to run

Agent Sidecar run

Auth refresh fix

guarded

Without Agent Sidecar

unchecked final

$2.84

per run

repeats auth callback edit
skips browser login flow
same sign-out bug returns

With Agent Sidecar

validated fix

$1.41

per run

adds task contract
blocks premature final
verifies login flow

Agent Sidecar runs beside Codex and Claude Code, corrects failure modes mid-run, and proves whether the intervention helped

Why Agent Sidecar

AI coding agents are useful, but uncontrolled runs get expensive

The warning

AI budgets can disappear fast

Fortune reported that Uber burned through its 2026 AI coding tools budget in four months after Claude Code adoption surged

Fortune on Uber

That is where Agent Sidecar comes in: a runtime layer for Codex and Claude Code that cuts repeated loops, stale context, and wasted tokens so builders spend less per successful task.

Capabilities

Six sidecar controls for cheaper agent runs

01

Runbook memory

Reuse approved local fixes as short runbook notes when the repo, task shape, and failure signal line up

situation / avoid / unlock

02

Repo map cache

Keep hash-checked notes about important files so the agent does not keep rereading code it already understood

path / symbols / validation

03

Trajectory watchdogs

Catch repeated attempts, stale evidence, drift, and low-progress runs before they burn another round of tokens

drift / stale / repeat

04

Tool-cost guardrails

Score commands, edits, fetches, and searches for risk or waste before the agent stacks up avoidable work

command / edit / fetch

05

Prompt shaping and compaction

Convert vague requests into scoped task briefs, carry validation evidence forward, and trim low-value history

brief / constraints / proof

06

Ensemble steering

Ask cheaper configured models for narrow guidance when Codex or Claude Code needs direction, then spend less overall

low-cost steerers

Demo

Agent Sidecar inside a Codex workspace

monitor: validation after edit
agent-sidecar-demo
Codex hook active

Phase 1 Hook setup

Install the sidecar, start the daemon, and connect Codex hooks

Terminalsetup

$ python -m pip install agent-sidecar

$ agent-sidecar daemon start

$ agent-sidecar hook install --provider codex

hooks ready

00:48
How to use

Run normally, review everything

The dashboard is the control plane and replay UI. The core product is still the runtime pipeline: observe real agent runs, intervene on common failure modes, and measure whether interventions improve cost per successful task.

1Install the sidecar package with pip
2Start the daemon from your terminal
3Configure Codex or Claude Code hooks
4Run the coding agent normally
5Review traces, decisions, monitors, saved patterns, and A/B analytics
Acme account dashboard

Live runs

Runtime control plane

daemon connected

guarded runs

18

decisions

43

saved lessons

9

total cost saved

$4,820

Estimated from guarded runs compared with matched baseline traces

tracedecision

auth-refresh-fix

healthy

require validation

billing-webhook

watching

inject lesson

worker-memory-leak

review

block repeat

selected trace

auth-refresh-fix

policy: validation after edit
lesson: refresh token callback
action: require continuation